What Happens When AI Makes So Many Decisions That We Forget How?

What Happens When AI Makes So Many Decisions That We Forget How?

A personal investigation into how AI is changing our reflexes, our tolerance for effort, and eventually the decisions we still call our own.The most important thing AI may be changing is not what we know. It is what we do in the first few seconds after we realize we do not know something. That sounds trivial, but I think it may be the beginning of a much larger behavioral shift. For most of my life, uncertainty triggered a small sequence of actions: remember who might know, search for the right words, open several sources, compare them, decide what seems plausible, and only then act. Increasingly, one new reflex compresses that sequence into a single gesture: ask the AI. A strange symptom. A recipe for whatever is left in the kitchen. The name of an actor I can almost remember. A confusing bill. A message I do not know how to phrase. A question about whether a laptop is actually worth the extra $300. These are not spectacular uses of artificial intelligence. That is precisely why I find them more interesting than the demos. Technologies change a society most deeply when they stop feeling like technologies and become habits. Search engines did that once. They gave us the phrase “just Google it,” and eventually we stopped hearing how strange that sentence was. AI is producing a new version: “just ask it.” The next phrase may be even more consequential: “just let it handle it.” This essay is about what happens in the space between those three sentences. Before “Just Google It,” Questions Had a Price For most of history, information had friction. You asked someone who might know. You checked a manual or an encyclopedia. You went to a library. You waited until you could speak to a doctor, mechanic, teacher, salesperson, or relative who had dealt with the same problem. Even when the answer existed, getting to it cost time, attention, and often a little embarrassment: you had to admit to another human being that you did not know. The internet collapsed much of that cost. Search engines taught an entire generation a new cognitive skill: translate the messy question in your head into keywords, scan results, ignore obvious junk, open several tabs, compare sources, and assemble an answer from fragments. We outsourced access to memory, but we still did much of the interpretation ourselves. A famous 2011 Science study by Betsy Sparrow, Jenny Liu, and Daniel Wegner found that when people expected information to remain available, they were less likely to remember the information itself and more likely to remember where it could be found. Later work on cognitive offloading described the broader pattern: humans routinely move mental work into notebooks, calendars, calculators, GPS, phones, and other parts of the environment. There is nothing inherently unhealthy about this. Civilization is, in part, a history of building tools so the brain does not have to carry everything alone. But Google still left one important burden with us. It gave us routes to information. It usually did not decide which route deserved our trust. That distinction is where generative AI begins to feel different. AI Is Changing the First Reflex, Not Just the Answer With conversational AI, I no longer have to think like a search engine. I can ask the question the way it exists in my head, incomplete and slightly confused. The system can ask a follow-up, infer what I am trying to accomplish, summarize competing possibilities, and turn scattered information into something that feels like an answer. That changes more than convenience. It changes the first move. Instead of entering an information landscape and navigating it, I hand the landscape to an intermediary. The difference is subtle because the screen still looks like a place where I type words. Psychologically, however, the task has shifted from retrieval toward delegation of interpretation. For readers who want the technical layer behind this shift, I have a separate explainer on how modern AI systems work. Real-world usage already looks surprisingly ordinary. A 2025 NBER analysis of consumer ChatGPT use found that Practical Guidance, Seeking Information, and Writing made up roughly 77% of conversations in its dataset, while non-work use accounted for more than 70% of messages. The public story of AI is often about coding, corporate productivity, and frontier capabilities. A large part of the private story is simply people trying to get through Tuesday. This is the first subconscious change I would watch: the destination of uncertainty. We used to send uncertainty outward to friends, books, forums, and search results. Now we increasingly send it to a single conversational system. Once that becomes automatic, the AI does not merely answer questions. It begins to occupy the position in our lives where questions go first. The Disappearance of Cognitive Friction I do not think the useful question is whether AI makes us “think less.” That framing is too crude. A calculator removes arithmetic effort but can free attention for a harder problem. GPS reduces route memorization but can make an unfamiliar city usable. The more interesting question is which kinds of mental friction disappear, and what those frictions were quietly doing for us. In a 2025 CHI study, Microsoft researchers surveyed 319 knowledge workers and collected 936 examples of generative-AI use. Higher confidence in AI was associated with less reported critical-thinking effort, but participants also described a change in the location of that effort: less work generating a first answer and more work verifying, integrating, and supervising the output. That is not the same as thought disappearing. It is thought being reorganized. Still, some mental work matters precisely because it is inefficient. Searching forces us to encounter disagreement. Comparing products can reveal that the feature we thought mattered actually does not. Writing a bad first draft can expose that we do not understand our own argument. Struggling to remember something can strengthen memory. A 2025 randomized study of undergraduates found lower long-term knowledge retention among students who used ChatGPT as a study aid than among those who used traditional methods, a result the authors linked to reduced effortful processing. One education study cannot be generalized to all everyday AI use, but it captures a principle psychologists have known for a long time: sometimes the work is part of the learning. The phrase I keep returning to is cognitive friction: the interval between wanting an answer and having one. Some friction is pointless bureaucracy. Some is the resistance that makes a skill, a preference, or a judgment become ours. AI is extraordinarily good at removing the interval. The danger is not that every removed step makes us weaker. It is that convenience can erase the distinction between friction that was waste and friction that was formative. The real shift may not be that AI thinks for us, but that searching, comparing, and deciding start to feel increasingly unnecessary. (Source: AI-generated for this article.)Is AI Addictive? The Drug Analogy Is Tempting — and Probably Too Simple It is easy to see why people reach for the language of addiction. The loop is unusually smooth. A question appears, the assistant responds immediately, and the response is often useful enough to justify another question. The system is available at midnight, it does not become irritated, and it does not make us feel foolish for asking something basic. Each successful interaction teaches the brain a very simple lesson: when uncertain, come back here. But “it feels rewarding” is not the same as “it is a drug.” Current evidence does not justify treating ordinary or even heavy AI use as equivalent to substance addiction. Researchers have begun building scales for problematic ChatGPT use, conversational-AI dependence, and generative-AI dependency. One 2025 validation program across 1,333 participants in the United States and Singapore identified dimensions including cognitive preoccupation, negative consequences, and withdrawal-like experiences. At the same time, researchers writing in Addictive Behaviors have warned that the field risks pathologizing a new everyday technology before there is robust evidence of impaired control, distress, and functional harm. I think that tension is more interesting than the headline “AI addiction.” There may be a large middle territory between healthy tool use and clinical addiction: dependency without obvious pathology. I depend on GPS, cloud calendars, and search in ways that would have looked extraordinary twenty years ago. If they vanish, I can still function, but I feel the missing scaffolding immediately. AI could become a deeper kind of scaffolding because it does not only remember for us. It can compare for us, explain for us, draft for us, reassure us, and eventually decide for us. The warning sign may not be hours spent staring at a chatbot. It may be a growing irritation with any task that requires us to proceed without one. A ten-minute search feels too slow. Writing the first paragraph ourselves feels unnecessary. Calling three repair companies feels absurd. Uncertainty itself begins to feel like a technical problem that should have been removed. That is why I would ask a different question: not “Am I addicted to AI?” but “Which abilities now feel unreasonably difficult when AI is unavailable?” The answer tells us more about dependence than screen time alone. Trust Will Arrive Through Boring Success I doubt most people will ever sit down and make a philosophical decision to trust artificial intelligence. Trust will arrive much more quietly. The recipe works. The cable recommendation is correct. The summary saves twenty minutes. The hotel shortlist is genuinely better than the one we would have built ourselves. The assistant catches a fee we missed. Nothing dramatic happens — and that is exactly how a default is formed. Research on algorithmic recommendations suggests that willingness to accept advice is strongly affected by perceived autonomy. In three experiments published in Computers in Human Behavior, people were more willing to accept algorithmic recommendations when they retained more choice, such as receiving several options rather than a single prescribed answer. Another 2024 study of purchasing delegation found that perceived decision autonomy significantly shaped willingness to let AI participate in buying decisions. We may accept machine judgment not when machines become perfect, but when the interface gives us just enough control to feel that the decision is still ours. There is a paradox here. The better the system becomes at making good recommendations, the less often we have a reason to challenge it. And the less often we challenge it, the harder it becomes to know whether our trust is calibrated or merely habitual. A tool can earn trust through accuracy and still train us into passivity if the cost of checking it becomes higher than the perceived benefit. When AI Chooses, It Also Chooses What We Get to See This is where the question becomes philosophical for me. We usually imagine AI choice as a final step: there are ten televisions, and the agent picks number four. But a recommendation system has already made a more invisible decision before that. It decided which ten televisions entered the room. Research on AI-assisted purchasing makes this point explicitly: recommendation systems narrow consideration sets and prioritize options before the consumer evaluates them. In other words, AI can shape not only the answer but the menu from which an answer becomes possible. If my agent shows me three hotels, I may feel fully autonomous while choosing among them. I may never see the twenty-seven hotels it excluded, or know which preference it assumed mattered most. This matters because preferences are not always fixed objects hidden inside us, waiting to be discovered. We often construct them while choosing. Looking at cameras may teach us that weight matters more than zoom. Comparing three neighborhoods may reveal that we care about walkability more than an extra room. A system that becomes extremely good at predicting what I usually prefer can save me enormous time. It can also make it easier for my past behavior to keep selecting my future behavior. That is one of the stranger possibilities of personal AI: convenience could turn yesterday’s self into tomorrow’s default. If my agent learns that I usually choose the quiet hotel, the reliable appliance, the familiar food, the cheapest morning flight, and the same kind of vacation, it may become excellent at protecting me from choices I would probably reject. It may also protect me from choices that would have changed me. I do not mean that an AI agent will imprison us in a recommendation bubble by design. The point is more ordinary. Every good personal assistant needs a model of the person it serves. The better that model becomes, the more important a new question becomes: should the agent optimize for who I have been, who I say I want to be, or who I might become if it occasionally lets something unexpected through? The Important Word Is Not Intelligence. It Is Permission. Today, most consumer AI still lives on the safer side of a line. It can recommend. We execute. An agent crosses that line by combining a model with tools, memory, accounts, and permissions so it can pursue a goal across several steps. The assistant does not merely tell me which action is sensible. It gets hands. The design problem then changes completely. The question is no longer only “Is the model smart enough?” It becomes: What may it purchase? What may it cancel? What may it send in my name? How much money may it spend without asking? What is reversible? What is visible to other people? How quickly can I take control back? A 2026 study of a general-purpose agent introduced the useful idea of delegation regret: people sometimes disliked an agent’s action not because the outcome was objectively bad, but because the system crossed a boundary they would have wanted to approve. The study was small — twenty university students — but its pattern is intuitively important. Participants calibrated trust by task, not by agent. They tolerated more autonomy for low-stakes advisory work and demanded more confirmation for actions that were externally visible or difficult to reverse. That suggests that “Do you trust AI?” may eventually sound as crude as “Do you trust the internet?” We do not need one global answer. We need layers of trust. I may trust an agent to reorder detergent, prepare a tax checklist, compare insurance policies, or move a meeting. That does not imply that I want it to send a resignation email, approve a medical procedure, or transfer $15,000. A recent HackerNoon essay makes a similar point from the system-design side: agentic AI changes who owns a decision, not merely how a workflow runs. AI becomes truly transformative when it stops giving us answers and starts quietly coordinating the actions that follow them. (Source: AI-generated for this article.)An Ordinary Tuesday in the Agent Era When I try to picture this future, I do not imagine a humanoid robot walking around the house. I imagine a normal weekday in which entire chains of tiny coordination disappear. The future becomes more believable when the examples are boring. You come home and the television is already on the wall You are at work and remember that the guest-room television finally died. You tell your agent: “Replace it. Around 55 inches. Good for movies. Under $800. Avoid models with recurring reliability complaints.” That is all. The agent checks current models, professional tests, owner complaints, the wall dimensions stored from a previous home project, the ports on your sound system, delivery times, return policies, and local installer availability. It knows from earlier purchases that you usually choose reliability over saving $100. Within the spending limit you already approved, it orders the television, schedules delivery, books the installer, sends the mounting details, stores the warranty, and updates the household calendar. You leave work. When you get home, the television is on the wall. The striking part is that almost none of the individual capabilities are science fiction. Comparison sites, payments, delivery windows, calendars, contractor marketplaces, and product reviews already exist. The new thing is orchestration: one system carries intent across them without requiring the human to be the integration layer. The agent era begins when “Which TV should I buy?” becomes “Take care of the whole thing” — and by the time you get home, it is already done. (Source: AI-generated for this article.)The refrigerator stops becoming your problem A refrigerator starts making a strange sound. Today, that can become a miniature project: find the model number, search the noise, read forums, decide whether it matters, find a repair company, compare reviews, call during business hours, explain the issue, and negotiate an appointment. In an agentic version of the same life, you record ten seconds and say: “Figure out whether this needs attention.” The agent identifies the model from purchase history, checks the manual and known service bulletins, asks a safe diagnostic question, and — if service is warranted — finds an authorized technician who fits the calendar. It may be allowed to organize the visit freely but required to ask before any repair over $300. The intelligence is useful; the permission structure is what makes the intelligence livable. Your weekend changes before you have time to reorganize it Now imagine that on Friday morning the weather at your weekend destination deteriorates, the attraction you booked closes, and the hotel changes its cancellation policy. A future agent does not merely send three alerts. It notices that the original plan is collapsing, checks the rules you have given it, holds a refundable room somewhere else, proposes a revised route, and asks one meaningful question: “The original weekend no longer makes sense. I can switch everything to this alternative for $74 more. Want me to do it?” I have explored that travel-specific version of agentic delegation in more detail in this article, where the same shift from planning to acting becomes especially easy to see. That is a different kind of convenience. It does not reduce one search. It reduces the need to continuously monitor the world for changes that affect our plans. The agent becomes less like a search box and more like a background layer of attention. The Next Step After Delegation Is Anticipation The television example still begins with a command. But the most consequential future may begin when the command disappears. If an agent has enough memory, permissions, and context, it can notice needs before we articulate them: the warranty expires next month, the pantry is running low, the train is likely to be cancelled, the annual subscription quietly increased its price, the dentist has not been booked in a year. At first, anticipation will probably feel like excellent customer service. Later it could become the normal relationship between person and software. We will not ask the system to solve every small problem because it will already be watching for the conditions under which a problem tends to appear. This is where I think the psychological boundary becomes much more serious. Delegation says, “I have chosen a goal; please execute it.” Anticipation says, “Infer which goals deserve action from the pattern of my life.” The first transfers work. The second begins to transfer agenda-setting. An assistant that reminds me that my passport expires is easy to welcome. An assistant that decides I should replace my car because maintenance costs crossed a threshold is still understandable. But what happens when the same logic reaches softer parts of life? It notices I have not seen certain friends lately. It decides I am overworking. It learns which invitations I usually regret accepting. At what point does a helpful model of my preferences become an invisible editor of my days? Delegation becomes a new default when one instruction can trigger research, booking, purchasing, scheduling, and delivery without us managing every step. (Source: AI-generated for this article.) Delegation Debt: What Happens When We Forget How to Take Back the Task? There is another risk that is less dramatic than a rogue agent and probably more plausible: delegation debt. We delegate a task because the machine does it well. Then we stop practicing the task. Eventually the cost of taking it back rises, which gives us another reason to keep delegating it. We already know this feeling in simpler technologies. Many people can navigate a city far less confidently without GPS than they once could. That is not necessarily a tragedy; navigation skill may be a rational thing to trade for time. But AI can reach skills that are harder to classify as disposable: comparing evidence, drafting an argument, negotiating a purchase, recognizing when an explanation sounds wrong, or even deciding what information is missing before a decision can be made. This creates a strange loop. The more competent the agent becomes, the less competence I need to exercise. But the less competence I exercise, the harder it may become to judge when the agent is wrong. Trust then becomes partly self-reinforcing: I rely on the system because it is good, and I have fewer independent reasons to know how good it is because I rely on the system. That does not mean we should preserve every obsolete skill out of nostalgia. I do not want to memorize phone numbers just to prove I can. The more useful principle may be to preserve the skills required to supervise the things we delegate. If AI writes, we still need enough writing judgment to know when the result is hollow. If AI compares products, we still need to understand which criteria matter to us. If AI acts, we need to retain the ability to recognize a boundary it should not have crossed. What Do We Actually Want to Keep Choosing? The obvious response to all of this is to defend human choice. But I am not convinced that every choice deserves defending. Comparing twelve washing machines is not a sacred exercise in free will. Modern life contains enormous quantities of low-value administrative thought: passwords, forms, appointment windows, renewals, return policies, price comparisons, delivery coordination, and inbox triage. If AI removes that burden, it can increase human agency rather than diminish it. The difficult part is that low-value choices and identity-forming choices are not separated by a clean line. Buying socks is probably trivial. Choosing where to live is not. But between them sits most of everyday life: what we eat, how we spend weekends, which news we encounter, which hobbies we continue, which purchases we postpone, which friends we make time for. Repeated small decisions are part of the texture from which a life is built. So perhaps the real design problem is not maximizing automation. It is deciding where friction is worth preserving. We may want almost zero friction around changing a delivery time and considerable friction around sending a message that could end a relationship. We may want an agent to prepare three investment options but not choose one. We may want it to draft a medical question but never decide that a symptom does not require a doctor. I suspect good personal agents will eventually need something like a moral map of permission: not only what they can technically do, but what kinds of decisions the user wants to remain consciously present for. The machine will need a model not just of our preferences, but of our preferred relationship to our own preferences. Will We Trust AI? I Think That’s the Wrong Question I think we will trust AI in many domains, but that answer is almost meaningless on its own. We already trust machines constantly: elevators, card networks, spam filters, navigation software, recommendation systems. What matters is not abstract trust. It is calibrated trust — trust matched to stakes, reversibility, visibility, evidence, and the cost of failure. That is why the healthiest future may not look like an agent that never asks permission or one that asks permission for everything. Constant confirmation can become its own failure mode; when every action produces another approval box, attention turns into ritual clicking. Anthropic reported in 2026 that experienced Claude Code users enabled broad auto-approval more often while also interrupting the agent more often, suggesting a shift from approving every microscopic step toward intervening when something actually looked wrong. In a separate engineering report, Anthropic noted that users approved 93% of permission prompts in Claude Code, illustrating how repeated confirmations can create approval fatigue. HackerNoon contributors have made a related argument about human-on-the-loop governance: if every action requires approval, supervision can degrade into ritual rubber-stamping rather than meaningful control. Consumer agents will need their own version of that balance. The safest relationship may be neither obedience nor independence, but negotiated autonomy: act freely inside well-understood boundaries, slow down around irreversible actions, expose uncertainty when it matters, and make it easy for the human to inspect and reverse what happened. And this returns me to the psychological question. We may think we are deciding how much to trust AI, while the more important process is happening in reverse: repeated convenience is teaching us what no longer feels worth deciding ourselves. The Quiet Change I Think We Should Watch There was a time when “just Google it” sounded like a remarkable promise. Then it became infrastructure. We stopped noticing that it had changed memory, social habits, expertise, shopping, travel, and even the way an unanswered question feels. AI is crossing the same threshold, but the shift is deeper: search changed how we reach information; generative AI changes how much interpretation happens before the information reaches us; agents will change how much action happens after the answer. What unsettles me is not the idea that AI will suddenly take control of everyday life. It is the opposite: that it will become good enough at thousands of tiny tasks that handing them over feels obviously rational, one by one. No dramatic surrender. No single day when we decide to stop choosing. Just a growing collection of decisions that no longer seem worth the effort. Some of that will be liberation. I would happily live with less scheduling, less paperwork, fewer comparison tables, and fewer hours spent coordinating services that should have been simple. But not all friction is waste. Sometimes friction is where judgment is built, where a preference becomes visible, or where we notice that the thing we usually choose is no longer the thing we actually want. So the question I keep coming back to is not “How intelligent will AI become?” It is more personal: which parts of my life do I still want to be present for, even when a machine could handle them better? I am not sure I know where I would draw that line yet. That uncertainty may be the point — we should probably decide it before convenience decides it for us. If I let AI make decisions for me, am I also giving it the power to shape who I become — or to prevent me from becoming someone different?

Original Source

Read the full article at Hackernoon →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.