We say production is no longer scarce, but we still keep looking at the product
One of the strange things about the current AI conversation is that many people have already accepted the premise, yet have not changed the frame. They will say, quite reasonably, that production is no longer the hard part. Writing a draft, shaping a proposal, making a prototype, generating visual directions, assembling a small system: more and more of what used to require specialized skill, time, and coordination can now be pushed into a presentable form quickly.
And yet the conversation still circles around the product. Is the output valuable? Is it well packaged? Is it differentiated? Does it have a moat? We acknowledge that production has become easier, but we continue to ask questions as if the product itself were still the main site of scarcity.
That mismatch matters. When products become easier to produce, the thing worth examining is not only the product, but the logic that brought it into being. Why should this proposal exist? Why is this direction worth pursuing? Does this polished thing actually understand the problem, or has it merely been organized until it resembles an answer?
In the past, an idea had to travel a long way before it could be taken seriously. You had to write it down, sketch it, prototype it, gather materials, persuade other people to spend time on it. That process filtered out many ideas that were not clear enough, strong enough, or worth the effort. Not every thought made it to the table. Many were stopped by cost before they became visible.
That filter has become thinner. A thought can become a proposal quickly. A vague direction can turn into several complete-looking versions. An idea that has not yet been properly examined can acquire structure, titles, and packaging almost immediately. So the question is no longer, or not only, can this be made? It becomes earlier and more direct: should this exist at all?
This may be one of the first real differences AI begins to amplify. Tool fluency matters, of course, but it is unlikely to remain the deepest divide for long. The deeper divide is whether a person can recognize, while a possibility is still half-formed and before it has been dressed up as certainty, whether it is an opportunity, a distraction, or merely a beautifully organized illusion.
The hardest thing to identify is what looks right
AI does not usually mislead us most effectively by being obviously wrong. Wild factual errors, broken logic, proposals that cannot survive first contact with reality: these are annoying, but they are not the hardest problems. The more difficult material is the work that appears to hold together. It has structure. It sounds confident. The concepts connect. It may even read more smoothly than something a person produced under pressure.
This kind of output is not necessarily wrong enough to reject immediately. But it is not necessarily right enough to trust. It occupies a dangerous middle ground: plausible, but untested; fluent, but not yet rubbed against reality. AI is very good at producing this kind of seeming correctness, and people are very susceptible to fluency. When something has headings, layers, causes, and conclusions, we tend to assume that thinking has taken place. But looking thought-through is not the same as having been thought through.
When information was scarce, much of the work lay in finding materials, filling gaps in knowledge, and building a frame from which analysis could even begin. Now there are too many materials, too many frames, too many versions, and each version can be made to look respectable. The cost has shifted from getting something to deciding which thing is actually useful. That is harder work, because it does not add to a blank page. It filters an excess.
This is why discernment feels tiring. Production gives immediate feedback. You can see the pile growing. Discernment often looks like deleting, doubting, postponing, refusing, or starting over. It does not generate the same excitement. But it is what keeps excitement from becoming misdirection. In the AI era, the scarce skill is not spotting what is obviously wrong. It is spotting what is too smoothly right.
The faster things change, the more foresight matters
This does not mean everything should be figured out in advance. AI’s capabilities, interfaces, and application layer are changing too quickly for that. Many problems only become visible once you begin working. Sitting still until every possible direction has been clarified usually becomes its own form of avoidance. The more practical stance is to let vision, anticipation, and execution run together: to act while keeping some awareness of the longer cycle.
The point is not whether to act, but whether there is a stable enough direction to move through change. In the past, execution cycles were long. A rough idea passed through discussion, scheduling, resource constraints, technical implementation, and collaboration before it became real. These stages were slow, but they also created buffers. If the direction was weak, it often became visible somewhere along the way. It could be corrected, challenged, or stopped.
Now those middle layers are thinner. More importantly, the environment itself keeps being rewritten. What looks advanced today may become a default feature in a few months. What looks like a scarce capability may be absorbed by the model. An advantage built around a particular tool may lose its meaning after the next major iteration. The faster things change, the less useful it is to hand your attention to every local update. You have to distinguish between short-term noise and changes that actually alter the underlying problem; between what needs a response and what merely needs to be watched.
That is what foresight is for. It is not conservatism, and it is not a demand to redesign your direction every time the outside world moves. It is closer to vision: not predicting what each model release will do, but forming a relatively stable expectation about what is likely to be temporary, what deserves long-term investment, which capabilities tools will absorb, and which problems will remain human. AI has not reduced the need for thinking. It has simply moved more of that thinking forward.
In collaboration, the cost moves to verification
This shift becomes even more visible inside teams. In the past, much of the cost of collaboration was tied to production itself: who would write the first version, organize the materials, make the deck, create the mockup, turn a loose thought into something other people could discuss. AI lowers the cost of those acts dramatically. One person can generate ten versions quickly. Another can just as quickly produce an alternative. On the surface, collaboration becomes faster.
But the cost has not disappeared. It has moved. The new cost sits in verification. When everyone can quickly produce materials that look reasonable, the real work is no longer aligning around the existence of a deliverable. It is aligning around whether the deliverable should be trusted. What does this proposal assume? Is it solving the problem, or only the visible surface of the problem? Which option is merely well packaged, and which is actually worth pursuing? Is this conclusion supported by evidence, or held together by language?
AI makes team consensus more delicate because more of the material teams discuss is generated or assisted by AI. That material is naturally cleaner, more complete, and more finished-looking. Precisely because it looks complete, teams may agree too early. Everyone feeling that something “seems fine” is not the same as it being fine.
The future of high-quality collaboration will not only depend on whether everyone can produce more. It will depend on whether teams can discern together. What can be accepted quickly? What has to return to facts? What is only rhetorically coherent? What needs users, data, constraints, or time before it deserves confidence? The faster production gets, the more important verification becomes. Speed without verification simply distributes error more efficiently.
Clarity will be amplified, so will confusion
All of this points back to the same claim: discernment is harder than production. It is harder because it moves against our appetite for expansion. Production lets us see more things coming into existence. Discernment forces us to admit limits. It tells us that many things do not need to exist. Production looks like progress. Discernment can look like hesitation. But in an environment of excessive possibility, the ability to pause is not weakness. It is part of the work.
It is also harder because it is difficult to fake. In the past, a person with unclear thinking could be protected for a long time by long execution cycles, complex organizations, other people’s compensation, and some luck. If the direction was vague, there were still many stages to soften the consequences. If the thinking was imprecise, the result did not appear immediately. Once AI thins those buffers, the distance between thought and output gets shorter. What you think can become something quickly. What you have not thought through can also become something quickly.
This is the real divide. It will not remain, at least not for long, at the level of tool use. Tools will become easier. Production capacity will become more common. The difference will move downward: who is clearer about what they are doing, who can recognize errors that look reasonable, who can subtract when possibilities multiply, who can make speed serve discernment instead of replacing it.
AI will amplify clear people, and it will amplify unclear people too. The difference is that one amplification becomes capability, while the other becomes consequence. AI does not automatically make a person clearer. It only reveals, faster than before, whether clarity was there.
The question returns to choices
So perhaps the most important question in the AI era is not whether an output is complete enough, but whether the choices behind it hold. This question returns at every meaningful moment: when starting a project, accepting a direction, generating a set of options, or deciding to commit more resources. It is not abstract. It has to be answered through choices: what to do and what not to do; what to believe and what to question; what to continue and what to stop.
Discernment has no final state. It is not something you settle once and then leave behind. It is a continuous calibration. You act, test, revise, and adjust within the amount of load you can actually carry. Each cycle should make the original problem a little clearer. Discernment does not mean knowing the right answer in advance. It means retaining contact with the real question while possibilities keep arriving.
There will be more possibilities. The world will keep rewarding the impulse to do something immediately. But the more that happens, the more expensive discernment becomes. In the end, what determines how far you can go is not how much you can generate. It is whether, among all the things that seem to make sense, you can tell which one is actually worth it.
About the author
Qinqin (Stella) Yang is a creative AI builder and experiential creative director. She builds workflows, systems, and creative direction that operate under real constraints, and designs speculative frameworks that expand what teams can imagine before they start making. Over the past decade she has built brand experiences, product design, and speculative projects for Tencent, IKEA SPACE10, Burberry, and the Palace Museum. She is the founder of Actual Fiction, where she develops AI-native tools that turn imagination into a working method.
This newsletter is where she thinks out loud about creativity, AI, and what it means to build toward a future that doesn’t exist yet.
More at qinqinyang.com · Practice at Actual Fiction