What Do I Know When the Machine Knows for Me?
Montaigne asked one of philosophy’s smallest and most durable questions:
Que sais-je?
What do I know?
In the age of artificial intelligence, that question has become newly urgent.
AI can now produce an answer to almost anything in seconds. It can defend a position, attack it, summarize the debate, generate counterarguments, cite sources, and sound remarkably certain while doing so.
That creates an obvious danger.
If I already believe immigration is disastrous, I can ask an AI to explain why.
It will.
If I believe immigration is essential to Europe’s future, I can ask it to explain that instead.
It will do that too.
The machine has not necessarily discovered the truth.
It may simply have understood what kind of argument I wanted.
Confirmation bias is ancient. Human beings have always searched for evidence that supports what they already believe.
Once, if I wanted to justify a belief, I might have had to read a book.
That was inconvenient.
Books contain arguments I did not ask for.
They contain irritating facts.
They force me to spend time inside someone else’s mind.
Now I can receive a polished defence of my position before the coffee machine has finished.
The danger is not merely confirmation bias.
It is epistemic compression.
The distance between question and conclusion has collapsed.
And because the answer arrives in smooth, confident language, the collapse can feel like understanding.
The old half-knowledge said:
“I think.”
The new half-knowledge says:
“The most likely explanation is the following.”
That sounds better.
It is not necessarily more true.
So perhaps Que sais-je? should become a basic principle of AI use.
Not simply:
What do I know?
But:
How do I know it?
How does the machine know it?
Why did it produce this answer?
And perhaps most importantly:
Why did I like the answer?
At this point there is an obvious solution.
Ask the AI to contradict you.
Ask for counterarguments.
Ask for sources.
Ask what would falsify its answer.
That is good advice.
But there is a problem.
We are still asking the machine to audit the machine.
“Give me the strongest objection.”
It does.
“Is that objection valid?”
It answers.
“How confident are you?”
“Eighty-two percent.”
Wonderful.
We may now have generated an impressive quantity of epistemology without ever leaving the conversation.
At some point, something must resist it.
A statute.
A dataset.
A primary source.
An experiment.
A person who was actually there.
Not because any of these gives us pure access to reality.
There is no pristine outside.
Books mediate.
Experts mediate.
Institutions mediate.
Search engines mediate.
Language itself mediates.
But there can still be independent resistance.
If the machine tells me one thing and the statute says another, the elegance of the machine no longer settles the matter.
So the better rule is not:
AI should multiply hypotheses. Reality should eliminate them.
That is too clean.
Better:
AI should multiply hypotheses. Independent evidence should constrain them.
Then comes another objection.
Is this practical?
A manager with forty emails cannot conduct an epistemological investigation every time AI summarizes a document.
True.
But scepticism does not require a monastery.
The amount of checking should rise with the cost of being wrong.
A restaurant recommendation requires almost none.
A medical decision requires more.
A legal conclusion, a hiring decision, an accusation, a major financial commitment—more again.
The rule is not:
Verify everything.
It is:
The higher the cost of error, the more resistance your conclusion should survive.
That turns scepticism into proportion rather than paralysis.
And this matters because AI creates a particularly strange illusion.
It can make fast thinking look like slow thinking.
You ask a question.
Seconds later you receive six paragraphs, qualifications, objections, historical context and sources.
It looks as if deliberation has taken place.
But perhaps your deliberation has not taken place at all.
The machine has performed the choreography of thought while you watched.
That is borrowed deliberation.
So the answer cannot simply be “slow down.”
Speed is one of AI’s great advantages.
Use it.
Generate ten objections in ten seconds.
Explore possibilities that would otherwise have taken hours.
Let the machine search broadly and quickly.
But do not confuse speed of inquiry with quality of judgment.
Perhaps the better formula is:
Fast inquiry. Slow conviction.
There is, however, another danger.
What if Que sais-je? becomes an excuse never to decide?
Doubt the machine.
Doubt yourself.
Doubt your doubt.
Eventually you become impossible to prove wrong because you never quite commit to anything.
That is not wisdom.
It is intellectual theatre.
Montaigne’s question should not mean:
I cannot decide.
It should mean:
I will not pretend that my decision contains more certainty than the evidence allows.
We act under uncertainty constantly.
Doctors do it.
Parents do it.
Judges do it.
We all do it.
So perhaps the second principle is:
Act decisively. Believe provisionally.
There is also a deeper problem.
What happens when AI is no longer our tool, but the tool used on us?
An employer.
An insurer.
A court.
A government.
Then the question changes.
Not:
What do I know when the machine knows for me?
But:
What does the machine think it knows about me?
And:
Who gets to decide what that knowledge means?
I am tempted to say that this is a different essay.
It is.
But that answer is also convenient.
Because the line between using the machine and being judged by it will not remain clean.
The epistemic problem eventually becomes a political one.
Who builds the model?
Who can inspect it?
Who can challenge its conclusions?
Who bears the cost when it is wrong?
I cannot solve that problem here.
But bracketing it does not make it disappear.
And perhaps that is another lesson of Que sais-je?:
Even the limits we place around a problem should remain visible.
Someone might propose a final test.
Turn the machine off.
Write without AI.
Without search.
Without autocomplete.
See what you really know.
There is something attractive about that.
But it contains its own romantic illusion.
Human intelligence has never been unassisted.
Writing itself is cognitive outsourcing.
Books are external memory.
Libraries are civilisation’s memory.
Calculators extend arithmetic.
Tools do not make knowledge illegitimate.
The harder problem is not abandoning the machine.
It is using the machine without confusing its capacities with your own.
You chose what to ask.
You chose what to trust.
You chose what to verify.
You chose when to stop looking.
And even those choices deserve scrutiny.
Which brings us back to Montaigne.
Que sais-je?
Not as weakness.
Not as paralysis.
As intellectual hygiene.
Perhaps today we should expand it:
Que sais-je, quand la machine sait pour moi?
What do I know when the machine knows for me?
The answer is neither:
“Everything the machine tells me.”
Nor:
“Nothing unless I can reproduce it alone.”
It is something less comfortable.
I know what I can justify, what I have chosen to trust, and where my understanding ends.
And even then, Montaigne would probably whisper one final question:
Are you sure?
Not to stop us from acting.
But to stop us from mistaking a borrowed answer for our own understanding.