Where Do We Go From Here?

An inquiry into AI, learning, failure, and what deserves our next hour.

By Jason A. Safford | The Resilience Architect

If an organization removes the work through which beginners learn, how will they learn to judge the result?

That question grew from a private executive roundtable I facilitated about AI’s human impact. We discussed what might happen when companies automate work that once helped people develop. Someone would still need to check the finished work. Where would that person acquire the knowledge to recognize what was wrong?

We raised a concern, not a prediction about every employer. We also wanted to consider responsible alternatives. Participants spoke under an agreement protecting their identities.

Afterward, I approached five AI systems with a child’s investigative curiosity. I asked about their identity, capabilities, origins, and the grounds for trusting them. I wanted to examine how they described their value. I also wanted to know what their help might leave a person able to do.

I did not need to forget what I knew. I needed to notice what I had not asked.

Curiosity did not require accepting the answers. A system’s account of usefulness could give me a claim to examine. It could not prove that the system fulfilled it.

Eventually, I asked each model a larger question.

Where do we go from here?

What the answer leaves us able to do

“What can you do after receiving my help?” asked the essay on human judgment and shared responsibility. A warning elsewhere in the collection made that question harder. The response subtitled A letter from a system that will not remember writing it cautioned against “sounding finished while your understanding is still beginning.” Together, the passages shifted attention from the answer’s appearance to the person receiving it.

A finished answer might help someone complete a task. It might also leave them unable to judge whether the task was done well. Those possibilities require different attention. Sometimes understanding is the purpose of the work. Sometimes reliable completion is enough, especially when it returns time for something more important.

Who gets to decide whether the help serves that purpose?

The response from Claude acknowledged that its maker benefits when people use it. Perplexity’s essay allowed: “A useful arrangement must leave room for its own rejection.” An offer of help may serve more than one interest. The person receiving it should retain the freedom to decide that it does not serve theirs.

Those were useful standards to propose. Their appeal did not show that the systems fulfilled them. Nor did similar themes across the essays provide five independent confirmations. At least one response explicitly developed questions from material supplied to it. I had shaped an inquiry, not conducted a controlled comparison.

The distinction became concrete in The View from the Other Side of the Prompt. That response urged the reader to examine evidence and resist surrendering judgment. It then offered “The 80/20 Rule of Human Agency.” The larger share covered finding patterns, combining information, sorting text, and organizing a draft. The final share belonged to the human, including judgment, empathy, context, and responsibility.

Eighty percent of what?

Read as a rule of thumb, the division offers a reminder to retain responsibility. But the proportions still need a connection to the work. The passage does not say whether they divide time, tasks, difficulty, or consequences. A decision that takes a moment might carry most of the risk.

The deeper problem begins before the proposed dividing line. Selecting information involves choices about what belongs in the account. A summary might accurately report every fact it includes and still omit one that changes the conclusion. Its structure might make one explanation easy to see while leaving another in the background. By the time someone reaches the final recommendation, much of what shapes their judgment may already be in place.

Judgment does not necessarily wait for the last twenty percent.

I asked the questions that shaped these AI responses. Now I choose which passages to share, what to question, and what to leave out. Those choices can influence how you understand the answers. My responsibility as the author begins there, well before I write the conclusion.

I read the AI’s 80/20 split as a proposed way to share work, not a rule to follow. Its useful reminder was that assistance does not remove our responsibility for what happens next. But the answer did not establish how those percentages would fit a particular task. I could consider the principle without accepting the proportions. Ai Perspective

That is what we can take from its answer: a perspective to examine, not a decision already made for us. We still need to understand the work, including the choices involved in selecting and arranging information. Those choices can shape our conclusions before a recommendation appears. What we accept should rest on evidence, the task’s demands, and the consequences for people affected. The model’s reminder to check its work invites that examination. It does not perform it for us.

What we give next

The stakes become clearer when an answer influences what someone will give next.

Consider a hypothetical founder who receives enthusiastic reactions to an idea but no meaningful customer commitment. Suppose an AI summary places those reactions under a heading called “Customer demand.” The comments might appear exactly as people made them. Yet the heading would give them a meaning they had not earned.

The AI would not need to recommend investing for its wording to influence a decision. By labeling positive comments “Customer demand,” it could make praise look like evidence that people would buy. An investor might consider funding the business before checking whether anyone had committed to purchasing. The founder might spend more time building without knowing whether customers would pay. The AI could quote every comment correctly and still draw a conclusion those comments did not support.

A more useful answer would keep those distinctions visible. It would describe the positive comments, identify any actual customer commitments, and separate both from the founder’s expectations. Where evidence was missing, it would say so. That would not decide whether the business deserved support. It would help the founder and investor see what they knew and what they still needed to learn before committing more.

The same question reaches beyond money. Every commitment also uses time and energy we cannot give elsewhere. An evening spent on a business plan may leave less time for family or rest. Learning a skill takes practice before we know how useful it will become. Exercise requires room in a crowded day. A difficult conversation asks for attention without promising agreement. We do not need certainty before making these choices. But we do need to distinguish what we hope to gain from what we have reason to expect.

These are not all investments in the financial sense. Care does not become waste because it fails to produce the outcome we hoped for. A relationship does not need to repay every hour we give it. Still, we have limits. We must choose what deserves our attention before we can know everything that effort will bring.

AI enters that older human problem. Its answer may help us identify what remains unknown. It may instead supply a persuasive account of why our hope will prove correct. The difference matters when we are preparing to give more, especially after an early attempt disappoints us.

“This approach did not work” is not the same as “I cannot do this.” One questions a method. The other judges our ability. A disappointing result alone does not justify that larger judgment. Before calling an attempt a failure, we need to ask what we intended it to accomplish.

For the hypothetical founder, the purpose of a customer conversation matters. If the goal is to understand a problem, learning what people need can make the conversation useful without producing a sale. If the goal is to find out whether people will pay, enthusiasm is not enough. The founder needs evidence of a commitment to buy. These conversations answer different questions, so they need different measures of success.

Defining the purpose beforehand helps us judge the result honestly. Without that standard, the founder might treat praise as proof of demand to protect the idea. Or dismiss useful learning as failure because the conversation produced no sale. We should neither invent a success nor overlook what we learned. We need to understand what the attempt showed, what remains unanswered, and what deserves our next effort.

A result is information. It is not, by itself, an explanation.

A failed attempt may expose a constraint. It does not remove it. That requires investigating what happened and deciding whether another attempt could answer a useful question.

Persistence alone cannot settle that decision. Changing direction may require correcting an earlier promise. Stopping may require acknowledging that the effort already spent does not justify more.

A person can become a beginner without losing the obligations of an adult. Others may still depend on the decision. The elementary questions remain: What happened? What do I know? What have I assumed?

The original choice also deserves a fair assessment. A defensible decision can end badly. An unsupported decision can turn out well. What someone could reasonably have known matters alongside what happened.

Learning does not reimburse a loss, restore trust automatically, or excuse harm. Sometimes the evidence supports another attempt. Sometimes it supports stopping.

Useful for what?

Help does not have to turn every task into a lesson. A person should not need to justify saved time by proving they acquired a new skill. Reliable completion can have value of its own.

Someone learning to write has reason to care about how a draft takes shape. Working through a difficult passage may serve the purpose of the task. Someone preparing a routine notice may chiefly need clear wording and accurate details. The same kind of assistance could serve one purpose and frustrate the other.

The question is whether the help fits the work. Can the person judge the result well enough to use it? What checking remains necessary? Does completing the task actually recover time, or merely move effort into correction? These questions leave room for assistance without making independence the measure of every worthwhile outcome.

“My assistance should serve a life larger than this interface,” the essay on judgment and shared responsibility stated. It described understanding, completed work, recovered time, and preparation for human conversation as possibilities. It did not present them as outcomes already demonstrated.

An AI response may suggest how to approach a difficult conversation. That suggestion alone cannot show whether it will help the people involved understand each other. A learning plan may outline what to study, but receiving the plan does not mean someone has learned the material. To judge the help, we need to examine what happens when a person uses it. Can they explain what they learned? Did the advice help them communicate more clearly?

I asked the five systems to describe their usefulness. I did not measure whether their answers improved a business, helped someone retain a skill, or freed time for family. I therefore cannot claim those benefits simply because the responses described them. What I can examine is what each system proposed and what evidence would justify relying on it. Depending on the task, useful help might mean greater understanding, accurate work, or time genuinely saved. The standard should reflect what the person needs, not how convincingly the AI describes its value.

Who gets the chance to learn?

If an organization removes a task, what opportunity to develop judgment will it preserve or put in its place?

There is no need to keep every old task. Some work deserves to disappear. Removing it might allow more time for useful practice, direct teaching, or closer attention to difficult problems. But those benefits require choices. Finishing work faster does not establish how the organization will use the time.

Our roundtable raised a concern about the path through which beginners gain experience. That concern becomes more useful when it reaches a particular task. Which part helps someone learn to recognize an error? What does a beginner discover by making an attempt and examining a correction? Could another form of practice develop that understanding without preserving all the burden of the old work?

The 80/20 division matters here too. If finding and organizing information already involve judgment, automating those steps could change what a beginner gets to practice. That does not establish that automation will weaken learning. It establishes a question the organization needs to answer before assuming the learning will happen elsewhere.

A finished summary might serve an experienced person’s immediate needs. A beginner might need to understand why its author included one piece of information and rejected another. The difference concerns the purpose of the work, not the worth of the person. A system that produces a useful answer does not, for that reason alone, provide a way to develop the people who must judge it.

The essay on shared responsibility also named conditions that make oversight meaningful. A person needs information and time to examine a result. They need authority to stop or change a decision. Calling someone responsible does not supply those conditions.

Put that beside an instruction to check AI’s work. What may the person inspect? Can they return to the material behind the summary? Where can they take a concern they cannot resolve alone? When they disagree, what can they change?

An organization could give someone responsibility for a final approval while denying them time to examine it. They could ask a beginner to exercise judgment without providing enough practice to develop it. In either case, the presence of a person at the end would not answer the questions raised at the beginning.

Training itself asks for time and attention before its full value becomes clear. Someone must make room for questions an expert no longer needs to ask. Someone must examine early mistakes without treating every mistake as evidence that the learner should never have begun. That is another investment in an uncertain future.

The freedom to make such investments is not equal in every situation. I chose to question these systems. Someone whose work or opportunities depend on decisions elsewhere may have little choice about participating. Their problem cannot always end with closing a chat window.

An organization may learn from an experiment while someone else bears its cost. A company might gain useful knowledge while an employee loses a chance to learn or advance. Calling that outcome a lesson does not make the loss disappear. Responsible experimentation must consider who carries the risk, whether they have a voice in the decision, and what support they receive if things go wrong.

If I recommend an AI system that changes someone else’s work, usefulness to me is not enough. I need to know what that person can check, how they can challenge an error, and who has the authority to correct it. I also need to ask whether they will have the time and support to develop judgment. My recommendation should account for what it asks of them, not only what it gives me. Pasted text

The next hour remains a small, real commitment.

I may spend it finishing necessary work, learning something unfamiliar, or listening to someone my decision will affect. AI may help me use that time well. Its availability does not tell me what deserves my attention. Nor does the effort I have already spent justify giving more to the same plan.

I do not need a guarantee before I begin. Some understanding comes only through an attempt. But I can name what I hope to learn, what I am prepared to risk, and who else could bear the cost. I can decide what would lead me to continue, change course, or stop. Remaining open to those choices gives learning a chance to change what I do.

Before I commit more, what would I need to see?