Imagine opening your inbox and finding 10,000 mathematical proofs.
Not summaries.
Not guesses.
Full candidate proofs.
Some are wrong. Some are correct. Some are trivial variations of known ideas. A few may contain something genuinely new.
No human being has time to read them all.
For most of scientific history, one of the great limitations was the scarcity of answers.
What if AI reverses that problem?
What if answers become abundant, while attention, interpretation, and judgment remain scarce?
That possibility may change the role of the scientist more deeply than automation alone.
Prefer to watch? Continue with the video below, or keep reading.
The Old Bottleneck Was Production
Scientific work has traditionally been expensive in human time.
A mathematician may spend months finding one proof.
A chemist may spend years developing one reaction.
A physicist may devote a career to one difficult measurement.
The slow pace of knowledge production meant that human attention could roughly keep up with the flow of results.
AI can break that balance.
A model can generate many more hypotheses than a laboratory can test.
It can propose many more proofs than mathematicians can inspect.
It can search many more molecular candidates than chemists can synthesize.
The limiting resource moves.
We can write the shift roughly as
\[ \text{scarcity of candidate ideas} \rightarrow \text{scarcity of evaluation and understanding}. \]
Proof Is a Good Preview
Mathematics offers an unusually clear version of this future.
Suppose AI becomes very good at generating formal proofs.
A proof assistant can verify logical correctness.
Now one important part of the problem is automated:
\[ \text{AI proposes} \rightarrow \text{formal system verifies}. \]
At first this sounds like the end of the story.
But imagine the machine produces 100,000 correct theorems.
Which ones matter?
Which reveal a new structure?
Which connect previously separate areas?
Which deserve names?
Which should appear in textbooks?
Correctness does not answer these questions.
The scarce resource becomes mathematical taste.
A Result Can Be Correct and Still Be Uninteresting
Science is full of true statements that no one cares about.
A large enough system can generate endless correct consequences of known facts. Jorge Luis Borges imagined something similar in The Library of Babel: a library containing every possible book. Somewhere inside it are all true discoveries—but they are buried in an overwhelming ocean of meaningless combinations.
That is the problem with abundance.
Discovery is not measured by the number of true sentences produced.
Important results often do something more:
- resolve a meaningful uncertainty,
- reveal a mechanism,
- compress many observations,
- connect previously separate ideas,
- enable a new technology,
- create a new research direction.
These are judgments of significance.
They depend on context.
And context is difficult to reduce to a simple score.
Understanding Is Different from Verification
Suppose an AI produces a proof that has been formally checked.
We know the conclusion follows.
But a mathematician may still ask:
Why is this true?
That question sounds strange. The proof already tells us why.
But not all proofs explain equally well.
One proof may be thousands of steps long and reveal little structure.
Another may expose one simple idea that makes the theorem feel inevitable.
The second proof gives understanding.
The same distinction appears in science.
A predictive model may be correct.
A mechanism tells us why.
A useful theory tells us what else should be true.
AI may make verified answers abundant before it makes understanding abundant.
The Literature Problem Becomes Extreme
Researchers already struggle with the volume of scientific publishing.
No one can read everything in even a narrow field.
AI helps by searching, summarizing, and synthesizing.
But AI-generated research could make the problem much larger.
Imagine a world in which autonomous systems produce:
- thousands of candidate hypotheses each day,
- automated experiments around the clock,
- machine-written analyses,
- formal proofs,
- simulations,
- proposed mechanisms.
The scientific literature cannot simply expand without limit.
Some layer must filter, rank, compress, and connect the results.
This becomes a new scientific function.
The Scientist as Curator of Meaning
One possible future role for human scientists is not primarily generating candidates.
It is deciding what deserves to become part of the shared scientific picture.
That includes asking:
Is this result important?
Does it change what we believe?
Is it robust?
Does it connect to another field?
Is the explanation satisfying?
Should anyone spend resources following it?
This is not passive curation.
It is intellectual judgment.
The scientist becomes partly an editor of machine-generated possibility.
Problem Selection May Become More Important
If solving becomes cheaper, choosing what to solve becomes more valuable.
Suppose AI can attack almost any well-specified problem.
Then the advantage shifts toward people who can identify questions that are:
- important,
- neglected,
- tractable enough to learn from,
- connected to larger scientific goals.
A mediocre answer to a great question may matter more than a perfect answer to a trivial one.
This is why scientific taste may remain central even in highly automated research.
Scientists Also Decide What Counts as Evidence
AI systems can rank data, but evidence is never completely context-free.
A surprising result may be dismissed as noise.
Or it may trigger a new field.
The difference depends on judgment.
Researchers ask:
- Was the measurement reliable?
- Was the experiment well controlled?
- Is the effect large enough to matter?
- Could an unknown confounder explain it?
- Does the result deserve replication?
These decisions shape the scientific record.
If AI systems make them automatically, the criteria must be made explicit.
That is both technically difficult and philosophically important.
Human Values Enter Before Application
Scientific discovery is not the same as deciding how a discovery should be used.
Suppose AI discovers a powerful new material.
Should it be manufactured?
Suppose it finds a biological intervention with dual-use potential.
Should the experiment continue?
Suppose it optimizes a technology that is economically valuable but environmentally harmful.
Scientific capability does not determine social value.
Humans remain responsible for decisions that depend on ethics, priorities, and public consequences.
The more powerful automated discovery becomes, the more important this distinction becomes.
Could AI Also Learn Scientific Taste?
Perhaps.
Taste is not necessarily mystical.
It is partly learned from examples.
Scientists develop judgment by studying important results, failed approaches, elegant proofs, productive questions, and the history of their fields.
An AI system could learn similar patterns.
It might rank results by novelty, explanatory compression, downstream influence, or disagreement with current theory.
It may even become better than most humans at identifying promising directions.
But even then, the problem does not disappear.
Someone must decide what the ranking is for.
Novelty?
Utility?
Beauty?
Human welfare?
Commercial value?
Truth-seeking?
Different objectives produce different science.
Explanation May Become a Translation Problem
There is another possibility.
AI may discover structures that are valid but difficult for humans to understand directly.
Then part of the scientist's job may be translation.
The machine works in a high-dimensional representation.
The human asks for a simplified conceptual model.
The AI produces an approximation.
The scientist tests whether that approximation preserves the important structure.
Science becomes a dialogue between different forms of intelligence.
This could resemble the relationship between a complex numerical simulation and a simplified analytic model.
The full computation may be inaccessible to intuition.
But a reduced model gives us something we can reason with.
The Value of “Useless” Questions
There is a danger in making scientific AI too goal-directed.
If every project must optimize a metric, we may lose curiosity-driven exploration.
Some of the most important scientific ideas began without obvious practical value.
Pure mathematics is full of examples.
So are fundamental physics and astronomy.
Humans sometimes investigate a question simply because it is beautiful, strange, or intellectually irresistible.
An AI system trained only to maximize near-term utility might neglect such work.
Scientific culture has always protected some space for curiosity.
That may become even more important when machines make optimization extremely efficient.
The Human Role May Move Upstream
It is tempting to ask which scientific tasks AI will take over.
Perhaps a better picture is that humans move toward different parts of the process.
From calculation to problem selection.
From routine analysis to interpretation.
From generating candidates to evaluating significance.
From executing experiments to defining goals and constraints.
From proving that something works to understanding what it means.
This is not guaranteed.
AI may advance in all of these areas too.
But the structure of scientific labor is likely to shift.
When Answers Are Cheap, Questions Become Valuable
The deepest change may be simple.
For centuries, science was constrained by our inability to obtain answers.
AI may create a world in which candidate answers are plentiful.
If that happens, the central question changes from
Can we solve this?
to
Which problems deserve to be solved, which answers deserve to be believed, and which results deserve to shape our understanding of the world?
That is where human judgment may remain most important.
And it gives us a natural place to turn toward quantum science.
Quantum systems push nearly every challenge we have discussed—observation, verification, representation, control, and learning—to an extreme.
They may therefore be one of the best places to test what AI for Science can really become.
Next in the series: Why Quantum Science Is a Special Test for AI
0 comments