The First Plausible Answer Is Already a Design Decision

I opened a blank document and asked an AI assistant for an outline.

It returned nine sections, four principles, and a conclusion that sounded as if the matter had already been settled by a small but efficient committee.

Nothing was obviously wrong.

That was the interesting part.

Before I had decided what I thought, I had a structure that made one interpretation of the subject feel natural. The model had not merely saved me some typing. It had selected the question, arranged the priorities, and supplied the shape of a satisfying ending.

The model did not steal the problem. I put it in the prompt.

This is less dramatic than a robot uprising, but considerably more common.

AI often changes the point at which we encounter our own work. We ask for an architecture before identifying the awkward constraint. We request names before deciding what the thing should feel like. We ask for five ideas and receive one sensible idea wearing five jackets.

The issue is not that the answers are always poor. Frequently, they are useful.

The issue is that a plausible first answer has gravity. Once it exists, later thinking tends to become revision rather than discovery.

That is a design decision, even when nobody remembers making it.

The philosophical traditions in this article are not interchangeable ingredients in one universal theory. William James, Rambam, Confucius, the Laozi, the Zhuangzi, and early Buddhist texts were addressing different problems in different intellectual worlds.

I am using each as a separate diagnostic:

• What changes in practice?
• What has actually been demonstrated?
• Does the name correspond to the role?
• Has the learner participated in the learning?
• Is empty space enabling the tool?
• Is usefulness being judged from only one position?
• Is this the right answer to release now?

Together, they offer a practical way to use AI without either romanticising manual effort or treating every removable difficulty as a bug.

Not all friction is waste

Some friction is simply tax.

Nobody develops wisdom by manually renaming 400 JSON properties or copying data between two systems that should have agreed on a schema three years ago.

That work can go.

Other friction is formative.

It is the effort involved in naming a conflict, noticing that two requirements cannot both be satisfied, deciding which assumption is doing too much work, or sitting with an explanation that is correct but somehow incomplete.

AI can help with those activities too. The question is what remains ours after it helps.

Cognitive offloading is not new. Notes, maps, calculators, and search engines all allow us to restructure a task by placing some of its cognitive demand outside ourselves. Research does not support a simple story in which offloading is either beneficial or harmful. It can improve immediate performance while changing what is remembered, practised, or learned, depending on the task and how the aid is used.

A 2024 short-story experiment illustrates a related tension. Participants given access to AI-generated ideas produced stories that human evaluators rated more favourably on average, especially among writers with lower baseline creativity. Measured across the whole pool of stories, the AI-assisted ones also became more similar to one another. That is one bounded experiment, not a universal law of creativity. Still, it captures something familiar: an individual result can improve while the wider field becomes narrower.

The useful question is not:

“Did the tool remove work?”

It is:

“What kind of work disappeared, and which capacity was being exercised there?”

I am happy to offload obscure syntax, repetitive transformations, and the production of a first test fixture.

I am less eager to offload the decision about what the system is for.

The distinction will vary by person and task. A one-off result may be all I need. At other times, I need to retain the ability to explain, adapt, or challenge the result after the chat window has closed.

Renting cognition is sometimes perfectly sensible. I rent servers too. I simply prefer to know which capabilities I still own.

William Jamesfollow the change into practice

William James is often reduced to the claim that “whatever works is true”.

That makes pragmatism sound like a pitch deck.

James’s account is more demanding. He asks what concrete difference an idea makes in experience, and how it becomes verified, corroborated, corrected, or defeated as events unfold.

An idea does not become true merely because it is convenient or persuasive. Its consequences must survive contact with experience.

This is a useful correction to the way AI capabilities are discussed.

“This model reasons.”

“This agent understands the task.”

“This tool improves productivity.”

Each statement may be meaningful. None is meaningful enough on its own.

What changed after the tool entered the workflow?

Did it find defects that reviewers would otherwise miss?

Did it reduce total effort, or move effort from production into verification?

Did it improve the final decision, or merely the fluency of the first draft?

Did it expand the user’s range of options?

Did it make errors easier to detect, or easier to believe?

A coding assistant may produce a function in thirty seconds. If I then spend twenty minutes checking a dependency it invented, the work was not eliminated. It changed shape.

A writing assistant may make a paragraph more fluent. If the revised paragraph sounds more certain than its evidence permits, the surface improved while the thought became worse.

A research assistant may produce an excellent map of a subject. If I never open the sources, I may also have created a new distance between myself and the material.

James suggests evaluating the whole change in practice, including the habits that accumulate after repeated use.

After six months with the tool:

• Do I notice problems sooner?
• Do I ask better questions?
• Can I still begin without it?
• Am I more willing to test an idea?
• Has my range expanded, or has my work become more predictably competent?

The relevant principle is:

A capability claim is a prediction about consequences, not a description of how impressive the output looks.

One answer can be excellent and still create a poor long-term working relationship.

Rambamdo not let fluency upgrade the evidence

Rambam’s Moreh Nevukhim, the Guide of the Perplexed, was written in Judeo-Arabic within an intellectual world shaped by Torah, Aristotelian philosophy, and the medieval Arabic philosophical tradition.

Its intended reader is not simply anyone who feels uncertain.

The reader is grounded in Torah, trained in the philosophical sciences of the period, and troubled by apparent conflicts between serious intellectual and religious commitments. Equivocal language, figurative expressions, and parables are central to that perplexity.

This matters because the Guide is not a general celebration of remaining confused.

Perplexity begins when two commitments that deserve serious treatment appear unable to coexist. The task is to distinguish a real contradiction from one created by language, habit, or insufficient understanding.

In Part I, Chapter 31, Rambam describes limits on human intellectual apprehension. He also identifies reasons disputes persist: the subject may be unusually subtle, the inquirer may lack sufficient understanding, established habits may obstruct reconsideration, or the desire to win may have overtaken the search for truth.

The chapter does not say that every question is permanently open. Quite the opposite. Rambam sharply distinguishes matters established by demonstration from matters around which disagreement remains because demonstration is unavailable.

Part II, Chapter 25 makes that distinction operational. Rambam says that demonstrated divine incorporeality requires figurative interpretation of scriptural passages that appear corporeal. By contrast, the eternity of the world has not been demonstrated. He therefore refuses to reconstruct the text around a proposition that remains unproved.

That is a valuable discipline for AI:

Do not let fluency promote a proposition from plausible to demonstrated.

A polished demo is not a deployment result.

A benchmark is not proof of unrestricted capability.

A compelling explanation is not evidence that the explanation describes the system’s actual causal process.

A citation-shaped sentence is not a checked source.

The Guide‘s negative theology requires particular care here. Rambam is discussing what may properly be said about God. AI is not analogous to the divine, and negative theology is not a technical framework for machine learning.

The transferable method is much narrower. In some circumstances, a precise account of what cannot be established tells us more than an impressive positive noun.

Compare:

“The agent understands the repository.”

With:

“The system can inspect files available in its current context. It cannot establish that an omitted file is irrelevant, that its context is complete, or that a generated explanation matches the developer’s original intention.”

The second description sounds less exciting.

It is also more useful.

The final chapter of the Guide introduces another interpretive difficulty. Rambam first ranks intellectual perfection above bodily, material, and moral forms of perfection. He then closes by describing a perfected person’s conduct as directed towards hesed, tzedakah, and mishpat on earth.

Commentators disagree about exactly how intellectual and practical perfection relate in this conclusion. I will not settle that dispute in a product blog.

The modest point is enough: the final picture does not leave intellectual apprehension as a private slogan. It has consequences for how a person acts.

For AI work, I use four labels:

1. Supported: Directly backed by evidence the system or user can inspect.
2. Inferred: A conclusion drawn from that evidence.
3. Proposed: An option offered for consideration.
4. Unknown: Something the current material cannot establish.

Generated prose tends to make all four sound alike.

The labels restore differences that fluency hides.

Confuciusthe learner must return the other three corners

The Analects says (2.15):

學而不思則罔，思而不學則殆。

Translations vary, especially around the terms wang and dai. The central contrast is between learning without reflective assimilation and thought detached from disciplined learning.

Here, “learning” is not merely receiving information. In the wider Confucian context it includes study, practice, imitation, formation, and repeated engagement with a tradition. Reflection is not free-floating speculation. It works on material that has been learned and lived.

AI makes it easy to separate the two badly.

A system can supply large amounts of learned material without requiring the user to assimilate it.

It can also simulate reflection without being adequately grounded in material. Ask it to reason from weak or absent sources and it may produce a very articulate fog.

Analects 7.8 offers a more concrete pedagogical image. Confucius does not open the way for a learner who has not begun to struggle with the subject. He gives one corner and expects the learner to return with the other three.

This should not be turned into a claim that Confucius anticipated modern active-learning software.

The pedagogical pattern is still useful. A learner is not merely the destination of an explanation. The learner must do something with the first part.

Applied to AI:

• Ask the user for an initial framing before supplying a complete plan.
• Offer a hint before a finished solution when learning matters.
• Ask what the user predicts will happen.
• Present one worked example, then let the user complete the variation.
• Reveal the next layer after the user has made a meaningful attempt.

The principle is:

Do not make the assistant so complete that the learner never has to return another corner.

This is not appropriate for every task. Nobody needs a Socratic dialogue before resetting a password.

But when the purpose is to retain a skill or develop judgement, immediate completion can be a poor interface.

The Analects also contains the idea usually translated as the “rectification of names”, zhengming. In 13.3, Confucius connects correct names with coherent speech, executable affairs, and social order.

The passage is political and role-based. It is not a product-naming guide. Interpretations differ over the precise relationship between language, offices, social roles, and conduct, but reducing it to tidy terminology misses its force.

A name creates an expectation about how its bearer should act.

That makes the analogy to AI products useful, provided it remains an analogy.

“Assistant” suggests support.

“Copilot” suggests shared operation.

“Agent” suggests delegated initiative.

“Autonomous” suggests that meaningful intervention is no longer required.

These terms change how people behave around a system. They also affect which failures appear surprising.

A better name describes the actual role:

• A drafting tool that cannot send.
• A research aide that retrieves sources but does not verify every conclusion.
• A coding agent with write access to a development branch.
• A recommendation system that proposes options but does not select the objective.
• A summariser that may omit low-frequency or minority positions.

The principle is:

Name a system by the work it performs, the authority it receives, and the point at which a person takes over.

Correct naming does not guarantee correct use.

It makes misplaced expectations less convenient.

Laozithe empty part is part of the tool

Chapter 11 of the received Laozi considers a wheel, a vessel, and a room:

有之以為利，無之以為用。

The spokes, clay, and walls provide material form. The open hub, hollow vessel, and inhabitable space make their respective functions possible.

This is not a claim that absence is always superior to presence.

A wheel made entirely of empty space remains an ambitious transport strategy.

The clay matters. So does the hollow.

The point is relational. What is present and what has been left open work together.

Software development often treats an unfilled space as unfinished work.

A pause needs a suggestion.

A blank field needs a default.

A moment of uncertainty needs a recommendation.

A workflow needs one more automatic step.

Sometimes that is correct. Sometimes the missing part is what permits the user to act.

Useful forms of “empty space” in AI products might include:

• A place for the user to state their own framing before suggestions appear.
• An abstention path when the available evidence is insufficient.
• A preview before an irreversible action.
• A permission the system deliberately does not possess.
• A pause in which a human can revise the objective, not merely approve the route.
• An answer short enough to leave room for thought.

The principle is:

Treat space, delay, and non-action as parts of the design, not merely as missing features.

An abstention path is not the absence of product behaviour.

It is product behaviour for a condition in which speaking would be less useful than stopping.

Zhuangziusefulness moves

The Zhuangzi is a compilation containing several voices, literary strategies, and philosophical tendencies. Treating every passage as one systematic doctrine would be particularly un-Zhuangzian.

Its stories about usefulness are often simplified into the slogan that apparently useless things have hidden value.

That is partly right, but not enough.

In one story, a large tree survives because a woodcutter finds it useless.

Soon afterwards, a host kills a goose because it cannot cackle. Its uselessness does not save it.

When a disciple asks whether it is better to be useful or useless, Zhuangzi declines the simple choice. Even occupying a stable position between the two would leave a person trapped by the same fixed classification.

The lesson is not:

“Optimise for uselessness.”

It is closer to:

“Be cautious about treating usefulness as a permanent property.”

Usefulness depends on a purpose, a situation, a timescale, and a point of view.

A feature that appears unproductive this quarter may preserve flexibility needed next year.

A person learning slowly may look inefficient until the environment changes and they are the only one who understands how the parts fit together.

An unusual idea may perform poorly against today’s evaluation because the evaluation was designed around familiar possibilities.

Conversely, something praised as useful can invite overuse, dependence, or exhaustion.

The AI implication is not that optimisation is wrong. It is that one measure of utility can quietly become a prison.

Before optimising, ask:

• Useful for whom?
• Useful for what purpose?
• Useful over which period?
• What becomes fragile if this is used everywhere?
• Which value disappears because the metric cannot see it?
• What would this system look like from another participant’s position?

The principle is:

Measure utility across more than one timescale and from more than one standpoint.

Sometimes five generated options are not evidence of variety.

They are evidence that the same standpoint has become very productive.

A Buddhist release check

In the Abhayarajakumara Sutta, Majjhima Nikaya 58, speech is considered through several conditions.

Is it true?

Is it beneficial?

Is this the proper time to say it?

Whether the speech is pleasant is not decisive. Something agreeable but false or unbeneficial is not justified. Something difficult may be said when it is true, beneficial, and timely.

The discourse also places this judgement within compassion for living beings.

An AI system does not possess that intention merely because its interface sounds warm.

Tone can improve an interaction. It cannot establish that the model understands the person’s interests, the wider situation, or the consequence of its advice.

The modern analogy should therefore be applied to the combined system, including its designers, operators, and users.

Before releasing an answer:

1. Is the central claim adequately supported?
2. Is it useful for the task the person is actually trying to complete?
3. Is this the right moment, channel, and level of detail?
4. Would a question, warning, or pause be more helpful?
5. Does the tone create more confidence than the evidence deserves?

The principle is:

Treat every generated answer as a release decision, not merely as completed text.

A factually correct answer can still be badly timed, disproportionate, or disconnected from the real question.

A reassuring answer can still avoid the part that matters.

---

The Perplexity Loop

I now use a small loop before asking AI to move from exploration to recommendation.

It is less impressive than calling it a framework, which may be one of its strengths.

1. State the question yourself

Write one or two sentences before opening the tool.

They can be incomplete. They may be wrong.

The purpose is not to prove that you can work alone. It is to give your own attention the first move.

2. Select the role

Decide what the model is being asked to do:

• Retrieve.
• Transform.
• Explore.
• Challenge.
• Recommend.
• Execute.

“Help me with this” often blurs all six.

A model can be excellent at transforming material without being the right participant to choose the objective.

3. Separate epistemic states

Require a distinction between:

• Supported facts.
• Inferences.
• Proposals.
• Unknowns.

This will not guarantee that every classification is correct.

It makes the answer easier to inspect and harder to accept as one undifferentiated block of confidence.

4. Diverge before converging

Before asking for the best path, ask for materially different framings.

Not five phrasings.

Not five variants that share the same hidden assumption.

Different accounts of what the problem might be.

Here is the prompt I use when I want the model to widen the field without deciding for me:

MODE: EXPLORE, NOT DECIDE

Question:
My current framing:
Known constraints:
What I may be assuming:

Return:

1. Facts you can support from sources you actually accessed.
   Link each source. If you cannot access sources, say so and do
   not invent citations.

2. Inferences, clearly labelled as inferences.

3. Three materially different framings of the problem.
   For each framing, state its key assumption and likely failure mode.

4. Important unknowns and the evidence that would resolve them.

5. Choices that depend on purpose, taste, or values.
   Mark these as mine to make.

Do not recommend a final path.

End with the one question whose answer would most change the
available option space.

This is not a truth machine.

It is a role boundary.

The prompt makes it harder for the model to collapse retrieval, inference, and recommendation into one seamless response. It does not make unsupported claims supported, and it does not remove the need to inspect important sources.

5. Converge deliberately

Once the option space is visible, switch modes explicitly.

State the objective.

State which trade-offs matter.

State which assumptions you are accepting.

Then ask for a recommendation.

This preserves a distinction between discovering possibilities and selecting among them.

6. Keep selected unaided repetitions

For abilities you need to retain, occasionally perform the relevant step without assistance.

A developer who must debug production code should sometimes investigate before accepting a generated patch.

A writer who wants to develop a voice should sometimes produce the opening before requesting alternatives.

A learner who needs durable knowledge should sometimes attempt retrieval before searching.

The correct amount will vary.

The point is not abstinence. It is deliberate retention.

7. Run the residue test

After using AI, close the transcript and ask:

• Can I explain the result in three sentences?
• Can I adapt it if one assumption changes?
• Can I identify its weakest point?
• Can I distinguish evidence from inference and proposal?
• Do I know which choice was mine?
• Can I formulate the next question?

If nothing remains after the chat closes, the productivity may have been rented rather than acquired.

Again, renting is fine.

The test merely tells us which transaction occurred.

What should AI make us better at?

Most AI products are optimised for time to first output.

That makes sense. It is visible, measurable, and easy to demonstrate.

For tools intended to support thought, learning, or innovation, we might also care about:

• Time to independent understanding.
• Diversity of genuinely different options.
• Ability to revise the objective.
• Quality of the next question.
• Retention of important skills.
• Ease of detecting uncertainty.
• Ability to stop or reverse an action.
• Whether disagreement becomes easier, not merely slower.

These measures are harder.

Difficulty of measurement is not evidence of irrelevance.

As systems move from producing text to taking actions, the division of labour becomes more important.

The useful question is no longer whether a human remains vaguely “in the loop”.

It is whether the loop still contains meaningful human agency.

Can the person understand what is happening?

Can they change the objective?

Can they disagree before the action becomes difficult to reverse?

Can they learn from the process?

Can they decide that this part should not be automated?

The goal is not to maximise human effort.

It is not to maximise machine participation either.

It is to find a division of labour in which abundant machine generation supports scarce human capacities: purpose, interpretation, taste, dissent, responsibility, and the ability to change one’s mind.

Leave the question open for another minute

I still use AI to create outlines.

Sometimes I use it before writing anything.

But I now try to notice the moment when I am asking for an answer mainly because uncertainty has become mildly uncomfortable.

Occasionally, I wait.

I write three bad sentences.

I identify the disagreement I have not resolved.

I decide whether I need evidence, expansion, criticism, transformation, or execution.

Then I open the tool.

This is not a ritual appreciation of difficulty. There is no philosophical prize for doing repetitive work slowly.

It is a small attempt to encounter the question before meeting the completion.

AI can finish the sentence.

It can expose omissions, generate alternatives, and help us move through unfamiliar material.

It can also make the first plausible path feel like the natural one.

The difficult skill is not keeping human hands on every keystroke.

It is keeping human agency alive in the direction of travel.

Funny enough, one of the more advanced ways to use AI may be leaving the question open for another minute.

---

Further reading

• William James, Pragmatism and The Meaning of Truth.
• Rambam, Moreh Nevukhim: the opening remarks, Part I Chapter 31, Part II Chapter 25, and Part III Chapter 54. The Hebrew text is Samuel ibn Tibbon’s classical translation of a work originally written in Judeo-Arabic.
• Analects 2.15, 7.8, and 13.3 at the Chinese Text Project.
• Laozi 11 and the Zhuangzi chapter “The Tree on the Mountain”.
• Majjhima Nikaya 58, the Abhayarajakumara Sutta.
• Evan F. Risko and Sam J. Gilbert, “Cognitive Offloading”, Trends in Cognitive Sciences 20, no. 9 (2016).
• Anil R. Doshi and Oliver P. Hauser, “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content”, Science Advances 10, no. 28 (2024).

Comments & Reactions

Got a thought, a war story, or a “well, actually”? Sign in with GitHub and jump in.

Loading comments…