Scientific instruments have always changed what humans can know.
A telescope sees farther.
A microscope sees smaller.
A detector sees wavelengths our senses cannot perceive.
A computer examines quantities of data that no person could inspect manually.
These technologies extended observation while leaving one distinction mostly intact:
the scientist decided what to do next.
That distinction is beginning to blur.
Self-driving laboratories combine robotics, automated instrumentation and algorithmic decision-making. Instead of merely collecting measurements requested in advance, some systems can use earlier results to determine which experiment should follow. Recent reviews describe a progression from narrow laboratory automation toward platforms that propose, execute and interpret experiments with increasingly limited human intervention.
AI research systems are creating the same transition at the intellectual level.
Robin can move from literature to hypotheses, from experimental results to analysis and then to revised hypotheses. The AI Scientist demonstrates a much more closed loop in a computational environment.
This may seem like a quantitative improvement: automation now covers more steps.
But it contains a qualitative change.
An ordinary analytical tool transforms an input.
An experimental agent can alter which future evidence will exist.
If a system reads the current evidence, selects an experiment and causes that experiment to be performed, its output is no longer merely a description of the world.
It has participated in creating the next state of the scientific world.
This does not imply consciousness.
It does not imply understanding.
It does not imply that the machine has become a scientist in the human sense.
But it changes the relationship between computation and evidence.
The system enters the causal loop.
That raises a problem which becomes more important as autonomy increases: provenance.
When a human scientist chooses an experiment, scientific culture asks for methods, reasoning, laboratory records and reproducibility.
If an autonomous system chooses the experiment, we need analogous visibility.
Why did it select this hypothesis?
Which evidence influenced the decision?
Which alternatives were rejected?
Was an apparently surprising discovery the result of genuine inference, hidden training contamination, optimization quirks or a chain of poorly documented software decisions?
The 2026 review of self-driving laboratories places provenance-complete experimentation among the major requirements for the field's next stage. That requirement may prove as important as raw autonomy.
Because science does not merely need results.
It needs results whose path can be reconstructed.
There is an irony here.
AI may allow science to explore more possibilities than humans ever could.
But the faster exploration becomes, the more important it may become to preserve a comprehensible record of how knowledge was produced.
The future laboratory could therefore contain two parallel systems.
One generates hypotheses, operates instruments and reacts to evidence at machine speed.
The other records enough of that process that humans can still decide whether the resulting claim deserves belief.
Autonomy without provenance would create discoveries that are difficult to audit.
Provenance without autonomy would preserve today's limitations.
The important scientific architecture may require both.
The transformation is therefore not simply:
human scientist → AI scientist
It is closer to:
human + model + instrument + evidence → a new distributed system of scientific agency.
And that may eventually make the question “Who made this discovery?” much harder to answer than “Is this discovery true?”