Answer 42

The culture around artificial intelligence has quickly produced a new oracle. Word the request correctly and the machine supposedly hands over an answer that used to require an education, experience and a few sleepless nights.

Douglas Adams pushed that dream to absurdity back in “The Hitchhiker’s Guide to the Galaxy”. A supercomputer spent seven and a half million years calculating the answer to the ultimate question of life, the universe and everything. The answer was precise: 42. The trouble was that nobody knew the question any more.

Something similar happens with generative AI. A well-presented answer can arrive before the person has worked out which problem he is solving, which data matter and how to tell a success from a mistake.

AI is not a solver but an amplifier. It speeds up the search for options, drafts, explanations and objections. A strong question turns it into a strong instrument. A vague question usually gets back a smoother version of the same vagueness.

The model has knowledge drawn from an enormous mass of text, so it can name a concept, an author or a method the user has never heard of. But it has no secret hatch into your reality. About your particular family, project, illness, contract or conflict it knows only what you told it, what follows from the available data and what it managed to reconstruct plausibly.

So the AI’s advice often turns out to be a reflection of the frame you set. Describe a colleague as lazy and the system will help you manage a lazy colleague. Mention that he has three incompatible tasks at once and a different analysis appears. The machine is under no obligation to notice a fact the person never reported and never thought to check.

Someone without a map of the subject does not always know which question is missing. He may not know about another jurisdiction, a hidden variable, a technical limitation or a premise disproved long ago. Receiving a confident answer does not automatically make him more knowledgeable — it only gives him, for a while, a result that looks like the result of a knowledgeable person.

The difference already shows up in research. Reviews in education find that generative AI can improve the finished assignment without producing a comparable gain in learning, if the student handed the thinking itself over to the machine. Labour-market experiments likewise find that the tool can narrow the gap in performance on a particular task, while the differences in knowledge remain once the assistant is taken away.

This is how borrowed competence arises. It is useful — like a calculator, a translator or a satnav. The danger starts when access to the result is taken for command of the method.

For AI to amplify the work, it has to be given not only the question but the context:

• what result is needed and for whom;
• what is already known and how it is known;
• which constraints cannot be broken;
• what counts as an acceptable error;
• which sources are permitted;
• where the system should admit uncertainty rather than fill the gap.

It is more useful to consult it in some mode other than “give me the answer”. Sometimes the best prompts are aimed against your own confidence:

What assumptions did I make?

What important context might be missing?

What does the strongest objection look like?

Which alternative explanations fit the facts?

What would change the conclusion?

That gives the model a chance to get past the first story. But the alternative it offers still has to be judged by a person able to understand its meaning and its consequences.

Generative systems produce probable text, not certified truth. NIST lists confident fabrications, the dependence of the output on context and excessive user trust among their risks. So a name, a link, a quotation, a calculation or a legal rule needs checking against the primary source.

You can hand AI a draft. You can ask it to explain, object, translate, compare and find the weak spot. You cannot hand it the job of being the person who understands the problem and carries the consequences.

Answer 42 becomes useful only once there is a question. Framing the question, choosing the context and checking the answer stay part of the human work.

Borrowed competence shows up most clearly in documents. A person asks the model to draft a formal complaint, gets four paragraphs with the right turns of phrase and sends it. The text looks like a lawyer’s work right up to the question “why thirty days here and not ten”. At that point it turns out the deadline came from a general formulation and not from the specific contract, which nobody attached. The model did not make a mistake: it was not asked about something it was never told.

The line between an instrument and an oracle runs along responsibility. A calculator is rarely wrong, but the person who typed the numbers answers for the result. It is the same here, except that the numbers are typed by confident prose and there is a good deal more room for a typo.