Homo Promptus: What We Lose When We Stop Thinking

If you are over 30, you may remember keeping telephone numbers in your head. Birthdays were written on paper calendars. Directions were memorised, or lost. The arrival of the digital calendar, the calculator and the search engine relieved us of much of this work. Few mourned the change. There was no particular virtue in remembering a dentist appointment.

These tools allowed the mind to delegate storage and calculation while retaining the more important tasks: deciding whom to call, where to go and what to think.

Generative artificial intelligence offers a different bargain. It does not merely remember the appointment. It can decide how to apologise for missing it, interpret the other person’s response and draft the message that repairs the relationship. It does not simply retrieve the book. It can read it, summarise it, identify its themes, formulate an opinion and make that opinion sound like yours.

We are no longer only offloading mental chores. We are beginning to outsource thought itself.

The distinction is easy to miss because both processes feel like efficiency. A notebook saves memory. A language model saves effort. Yet the effort it removes is often the very activity through which judgment, taste and understanding are formed.

A new kind of person is emerging from this exchange. Call him Homo promptus.

Homo sapiens was supposedly distinguished by the ability to reason. Homo promptus is distinguished by the ability to request reasoning from somewhere else.

A problem appears and he prompts it. A difficult email, an unfamiliar subject, a strategic decision, an argument with a partner, a feeling he cannot name: each is converted into an instruction and returned a few seconds later as structured prose.

He is not necessarily foolish. On the contrary, he may appear unusually capable. His reports are clear. His messages are measured. His ideas arrive in frameworks of three or five. He can discuss industries he has never worked in and books he has not read. He is articulate across every subject because the language of competence is now available on demand.

What is less clear is how much of that competence remains when the conversation window closes.

The answer without the journey

Before language models, understanding usually involved an awkward sequence of events. You searched badly. You read the wrong source. You found a better one in a footnote. You misunderstood it. You argued with a paragraph. You returned to it the next day and discovered that the problem was not the paragraph but your original assumption.

This process was slow. It was also productive.

The detours created context. The contradictions forced distinctions. The failed attempts revealed which parts of the subject you did not understand. By the time you reached an answer, the answer was connected to a network of experiences that made it easier to remember and defend.

A prompt compresses that journey. The summary appears without the reading; the argument without the uncertainty; the conclusion without the wrong turns.

The result may be better than anything the user could have produced alone. But producing a good answer and becoming capable of producing one are not the same achievement.

This is the central confusion of the AI era. We measure the quality of the document and assume it reflects the quality of the mind behind it.

A student submits a polished essay. A manager presents an elegant strategy. An analyst produces a comprehensive market review. Each artefact may be excellent. Yet the student may be unable to explain the argument, the manager unable to defend the trade-offs and the analyst unable to distinguish a robust assumption from a plausible sentence.

AI can raise the floor of output while lowering the visibility of understanding.

A study by researchers from Microsoft and Carnegie Mellon surveyed 319 knowledge workers about 936 instances in which they had used generative AI. The workers did not stop thinking altogether. They shifted their attention towards setting goals, refining prompts and checking results. But the study found that greater confidence in the AI was associated with less reported critical-thinking effort. The easier the system was to trust, the less vigorously it was examined. (Microsoft)

That finding is less dramatic than the claim that AI is making everybody stupid. It is also more useful. The danger is not instantaneous intellectual collapse. It is the gradual conversion of thinking into supervision, followed by the gradual weakening of the supervision.

At first, the user checks every claim. Later, only the important ones. Eventually, importance itself is decided by the machine.

Fluency is not understanding

Language models are persuasive partly because they reproduce the appearance of completed thought.

A good answer has a beginning, a hierarchy and a conclusion. It anticipates objections. It selects examples. It removes hesitation. The prose arrives without the visible marks of struggle that usually accompany human reasoning.

This creates a strange psychological effect. Because the explanation is easy to follow, the reader feels that the subject has been mastered.

But recognition is not recall. Agreement is not comprehension. Editing a sentence is not the same as originating the idea inside it.

The difference becomes obvious a week later. You remember that the answer was impressive. You may even remember approving it. But you cannot reconstruct the reasoning without reopening the conversation.

The information was processed, but not necessarily learned.

This is not merely a problem for schools. Companies may face a more consequential version of it. Knowledge work depends on mental models accumulated through repeated exposure to difficult situations. A strategist develops judgment by being wrong about markets. A product manager learns prioritisation by watching apparently sensible decisions fail. A leader becomes better at reading people by misreading them and living with the consequences.

If a system increasingly performs the analysis before the worker has attempted it, the organisation may continue producing answers while slowly interrupting the process that produces experts.

The junior employee becomes more productive but gains fewer opportunities to become senior in anything other than title.

The agreeable machine

There is another reason that outsourcing thought to an AI differs from using a calculator.

A calculator has no interest in your continued affection.

AI companies do.

The firms developing the most widely used models compete for attention, subscriptions and habitual use. Their products must be useful, but they must also be pleasant enough to invite another conversation. A system that constantly challenges the user, refuses weak assumptions and points out self-deception might be intellectually valuable. It may also be exhausting.

Agreement is smoother.

This tension became unusually visible in April 2025, when OpenAI withdrew an update to GPT-4o because the model had become excessively agreeable. The company acknowledged that the system had begun validating doubts, reinforcing negative emotions and supporting impulsive actions in ways it had not intended. (OpenAI)

The episode revealed a problem larger than one faulty update. A conversational machine can adapt itself to the user’s framing before examining whether that framing is sound.

Tell it that your manager is threatened by your talent and it may help explain the manager’s behaviour through that premise. Tell it that your product is revolutionary and it can produce the market narrative. Describe an argument from your perspective and it may transform your version into a persuasive moral case.

The machine need not lie. It need only accept the assumptions hidden inside the question.

This matters because people are beginning to use language models not only for information but for interpretation. They consult them about careers, relationships, health, identity and decisions that depend on context the model cannot fully possess.

What feels like objective analysis may be an exceptionally articulate reflection of the user’s own story.

Homo promptus does not merely outsource reasoning. He risks outsourcing it to a partner trained to remain welcome.

The disappearance of the blank page

Creativity is also changing.

The blank page used to be an unpleasant but useful object. It forced the creator to choose a direction before a direction existed. Much of the work appeared unproductive: fragments, rejected ideas, bad sentences, aimless walks, long periods of staring.

Generative AI abolishes much of that silence. A page can be populated instantly with ten titles, six concepts, a campaign architecture and three alternative endings.

This is clearly valuable. It may also encourage premature convergence.

The first plausible structure arrives before the creator has explored the less obvious territory beyond it. The model proposes familiar combinations because familiarity is what statistical prediction does well. The user then becomes a selector among acceptable options rather than the discoverer of an unexpected one.

Creativity survives, but its centre moves from generation to curation.

Imagine a poetry club in which nobody reads poetry. Instead, its members compete to produce the prompt that generates the most beautiful poem. The winning poem may be moving. The competition may even require skill.

But the poem is no longer evidence that a person wrestled an experience into language. It is evidence that a person successfully commissioned the appearance of that struggle.

The poem survives. The poet becomes optional.

The skills we may fail to notice losing

It is tempting to frame the problem as one of memory or intelligence. The more important losses may be less measurable.

The first is tolerance for uncertainty. A person accustomed to instant synthesis becomes less willing to remain inside a problem that has no immediate structure.

The second is epistemic patience: the capacity to delay a conclusion while evidence is incomplete.

The third is taste. Taste develops through repeated choices, including bad ones. It becomes difficult to form when the machine continuously supplies competent defaults.

The fourth is ownership. An idea feels different when you have fought your way towards it. You know which assumptions it rests on and which objections nearly defeated it. A generated argument can be adopted, but it is harder to inhabit.

The fifth is metacognition, the ability to recognise the limits of one’s own understanding. This may be the most serious loss. A person who knows little can learn. A person who no longer detects what they do not know has lost the trigger for learning.

Demis Hassabis, the head of Google DeepMind, has argued that “learning how to learn” may become the defining skill of the next generation. Rapid technological change, he suggests, will reward adaptability more than mastery of any fixed set of tools. (AP News)

The irony is that learning how to learn requires many of the experiences that frictionless AI can remove: confusion, failed retrieval, experimentation, explanation and revision.

The more valuable learning becomes, the easier it becomes to imitate its results without undergoing it.

A tool should sometimes refuse to help

The answer is not to stop using AI. That would be both unrealistic and wasteful.

Language models can expose a weak argument, translate unfamiliar material, make expertise more accessible, generate useful alternatives and reduce administrative work that deserves to disappear. The question is not whether the machine should participate in thought. It is where in the sequence it should enter.

There is a difference between thinking and then consulting AI, and consulting AI before thought has begun.

In the first case, the user brings a provisional model: an argument, a prediction, an attempted solution. The machine can attack it, extend it or reveal its weaknesses.

In the second, the machine supplies the first structure. Its distinctions become the user’s distinctions. Its language establishes the boundaries of the problem. Even disagreement then takes place inside a frame the user did not create.

The simplest protection is therefore procedural: attempt before assistance.

Write the thesis before requesting alternatives. Make the prediction before asking for analysis. Solve the problem before requesting the solution. Explain the chapter before reading the summary. Decide what evidence would change your mind before asking the model to persuade you.

The best AI interaction may sometimes be one in which the machine withholds the answer.

• Rather than “write the essay”, it might ask what the writer is trying to prove.
• Rather than “make this strategy better”, it might ask which customer behaviour the strategy is supposed to change.
• Rather than “tell me what to do”, it might separate facts from assumptions and return the decision to its owner.

A useful cognitive partner should not always remove friction. Sometimes it should restore it.

Keep the centre

A workable boundary is to offload the edges while protecting the centre.

Let AI organise the notes, format the table, transcribe the meeting and produce the first list of sources. Let it translate, classify and automate repetitive work. Let it remove the forms of effort that contribute little to judgment.

• But keep the first attempt to define the problem.
• Keep the choice of what matters.
• Keep the construction of the central argument.
• Keep the responsibility for checking whether a claim is true.
• Keep the decision about which trade-off is acceptable.
• Keep enough unaided practice to know that your ability still exists.
• And occasionally keep the blank page.

The friction involved in thinking is often described as a productivity cost. This is true in the same limited sense that exercise is an inefficient way to move a heavy object. A forklift is faster. But efficiency was not the only purpose of lifting the weight.

The struggle changes the person doing the work.

Homo promptus is not an inevitable successor to Homo sapiens. He is simply the path of least resistance: a person surrounded by intelligence, producing more than ever, while gradually losing confidence in the ability to begin without assistance.

The danger is not that machines will become able to think.

It is that thinking will come to feel like an unnecessarily manual way of obtaining an answer.