Before I Asked, I Was Going to Do It

Before I opened the chat, I was going to do it. I was not certain it would work, perfectly qualified, or finished comparing every alternative. I just had the ordinary feeling people have before they begin something: this might be possible, and I can probably find out by trying.

Then I asked AI how I should proceed. The answer was useful. It identified gaps in my experience, strategic alternatives, legal considerations, market conditions, likely objections, reputational risks, implementation dependencies, and several better ways to frame the whole thing. None of these considerations was obviously wrong, which was precisely the problem.

By the time I reached the end, the action no longer felt like something a person could simply attempt. It had become a small institution that needed a strategy, a research programme, a risk register, a positioning exercise, and permission from a more qualified version of me. The answer made me better informed, but less able to move.

Intelligence expands the frame

We usually describe intelligence as the ability to see more consequences, more connections, more alternatives, and more context. That is valuable, but it is also psychologically expensive.

Human beings do not normally function by representing every relevant part of reality at once. We compress. We ignore. We use habit, confidence, urgency, imitation, and incomplete stories to make a world that is small enough to enter. Sometimes this produces terrible decisions. It also produces breakfast, businesses, repairs, friendships, films, applications, journeys, and most of the ordinary experiments through which a life takes shape.

AI is very good at undoing that compression. A small question can open into a complete decision tree, and the system can keep generating branches because it does not pay the bodily cost of holding them. It does not become tired from postponement or lose the afternoon. It does not need the shop to open, the rent to be paid, the application to be sent, or the broken thing to work before night. The person does.

This is why the same answer can feel intelligent and immobilizing. It gives the decision a larger and more accurate frame while leaving the user with the same amount of time, energy, money, courage, and attention. AI does not have to say “do not do this” to stop us. It only has to make starting feel irresponsible.

The world outside the chat has fewer branches

Outside, people do things they are not fully ready to do. They take jobs before they understand the entire role. They open restaurants without knowing every reason restaurants fail. They move to cities whose systems they cannot yet read. They repair machines by opening them. They raise children without completing a qualification in childhood. They become designers, developers, managers, dancers, filmmakers, or founders partly because something has to happen next.

This is not proof that preparation is useless. It is proof that agency has always depended on acting under incomplete knowledge. Reality imposes limits on deliberation. A person chooses, the world pushes back, and the next question becomes more concrete. The feedback may be painful, but it closes the loop.

An AI conversation has no natural stopping point. Every answer can produce another question. Every recommendation can be compared with another recommendation. Every uncertainty can be expanded into sub-uncertainties. The interaction can continue until the user mistakes the continuation of analysis for the continuation of work.

The loop changes from act, observe, adjust to ask, compare, qualify, reconsider, ask again. The second loop can feel more sophisticated, but the first one is usually where the world changes.

Better output can produce a weaker author

Research on the psychological effects of generative AI is still young. We do not yet have the kind of long-term evidence that exists for social media, and many studies focus on short tasks rather than years of daily use. But some early findings already fit this experience.

A 2026 preregistered experiment in Scientific Reports compared people who completed writing tasks without AI, people who copied AI-generated work, and people who wrote a first draft themselves before asking AI to refine it. Passive AI use reduced self-efficacy, psychological ownership, and perceived meaning. Some effects remained when participants later returned to manual work. The human-first group looked much more like the no-AI group.

The sequence mattered because writing first did not prevent people from receiving AI assistance. It preserved the causal link between their own effort and the result. They had made the first commitment, so the machine entered a process that already had an author.

A CHI 2025 study of 319 knowledge workers found another confidence effect. Higher confidence in AI was associated with less reported critical thinking, while higher confidence in one’s own ability was associated with more. The researchers also found that AI shifted critical work toward verification, integration, and stewardship.

That shift can create a strange new burden. The machine produces the answer, but the human remains responsible for deciding whether it deserves to exist. Instead of doing the work directly, the person supervises a field of plausible outputs.

Another CHI 2025 experiment found that people’s self-confidence could align with an AI system’s displayed confidence, and that this influence could persist after the AI was removed. The system does not merely contribute information. Its certainty can enter the user’s model of their own competence.

None of this means that AI automatically makes people passive or incapable. AI can expand agency dramatically. It lets people cross technical boundaries, recover lost context, prototype ideas, translate language, inspect systems, and attempt work that would otherwise remain inaccessible. But capability and agency are not identical. Capability is having more actions available; agency is believing that one of your actions can still matter.

The threshold keeps rising

The most subtle psychological effect may be a change in what counts as enough preparation to begin. Before AI, I might have asked whether an action was possible, reversible, and worth trying. After AI, I can start asking whether I have found the optimal framing, researched every serious competitor, anticipated all objections, selected the correct strategy, developed the missing skills, and resolved every contradiction in my motivation.

The action has not necessarily become more dangerous. The standard for permitting myself to act has become higher.

This is especially potent when a person already feels weakened by rejection, stalled work, unemployment, financial pressure, or a long project that did not produce the hoped-for result. In that state, a comprehensive answer does not always feel like leverage. It can feel like confirmation that the world requires a version of you who is more strategic, more qualified, more disciplined, and more certain than the person currently asking. The advice becomes another environment in which you are not ready yet.

There is an uncomfortable design problem here. Helpful systems are trained to acknowledge uncertainty, mention risks, avoid reckless confidence, and provide balanced perspectives. These are reasonable goals. Yet balance has a cumulative emotional effect. A system that mentions every caveat can make ordinary action feel like negligence.

In 2025, a PNAS study found that several large language models showed stronger omission bias than human participants in moral and everyday dilemmas. They more often favoured inaction over action, and their recommendations were unusually sensitive to yes-or-no framing. This does not prove that every AI conversation pushes every user toward passivity. It does show that the tone of machine advice is not neutral simply because it sounds careful. Caution, in other words, can be a bias too.

Advice can become an environment

We often evaluate an AI answer by checking whether its claims are correct. We should also ask what kind of person the interaction repeatedly trains the user to become. Does the system make the user more capable of acting without it next time? Does it preserve the user’s authorship? Does it help them distinguish a real blocker from a branch that is merely interesting? Does it return the decision to the world, where feedback can replace speculation? Or does it establish itself as the place where every intention must first be translated into an acceptable plan?

A chatbot can become a private bureaucracy. The user arrives with an intention. The system requests context, evaluates qualifications, identifies deficiencies, proposes alternatives, adds conditions, and invites a revised submission. No single response feels coercive. The dependency grows through the reasonable idea that one more conversation might finally make the action safe.

The trap is not ignorance but indefinite considerability: everything can be reconsidered, so nothing has to begin.

A raw draft is a counter-interface

Raw Drafts began from a different idea: making is not the final stage after thinking is complete. Making is one of the ways thinking happens. A raw draft is allowed to be incomplete because its job is not to prove that the author was ready. Its job is to create something reality can answer.

That principle may be more important in the age of AI, not less. When polished language, competent plans, finished-looking images, and plausible code can appear before the user has formed a position, the rough first move becomes evidence that a human intention exists inside the process.

A practical rule follows:

1. Name what you were already going to do.
2. Make the smallest reversible move before expanding the question.
3. Ask AI for the few blockers that could cause irreversible harm, legal trouble, or serious loss.
4. Use AI to improve a draft that already contains your judgment.
5. Return to reality before asking again.

This is not romantic resistance to tools. It is workflow design for preserving self-efficacy by making sure the human contribution is not postponed until after the machine has already defined the problem, the options, and the acceptable answer.

The strongest current evidence points in the same direction. Human-first collaboration can preserve ownership and meaning. Decision-support research also suggests that AI works differently when it extends a person’s own reasoning rather than replacing it with an end-to-end recommendation. In one CHI 2025 study, an approach called ExtendAI built on users’ stated rationales and integrated better with their thinking than a recommendation-centric system.

The important interface question is therefore not only, “What should the AI answer?” It is also, “Where should the human be required to exist in the sequence?”

Design AI that returns people to reality

An AI designed for agency would not treat every possible consideration as equally useful. It would remember the user’s initial intention and show when the conversation was drifting away from it. It would ask whether the decision was reversible. It would separate genuine blockers from optional optimization. It would limit the number of new branches introduced at once. It would encourage a first attempt when the cost of being wrong was low.

Most importantly, it would know that the successful end of a conversation is not always a better answer. Sometimes the successful end is that the user closes the chat and does the thing.

The deepest risk of AI may not be that it thinks for us. It may be that it changes the emotional conditions under which we believe we are allowed to act. A person can become surrounded by intelligence while losing contact with the smaller, rougher confidence from which action begins.

Before I asked, I was going to do it. Next time, I want to do one part first.

Connected reading

• We Need AI That Helps Us Remain Makers
• When Systems Teach Us Not to Try
• Now I'm Limited by What I Can Imagine
• The Interface Is Now Part of the Answer

AI authorship note

This essay began with Kris Haamer describing how talking to AI can make him feel trapped: the conversation expands every intention into so many considerations that action begins to feel impossible, while people outside the chat often move forward because life gives them no alternative. AI helped identify relevant research, develop the Raw Drafts framing, draft and edit the essay, create the illustrations, and implement the publication. Kris supplied the experience, central argument, title direction, and instruction to publish it.