A common criticism of artificial intelligence is that it can summarize, rewrite, combine, and reorganize existing material, but it cannot discover anything genuinely new.

That criticism is partly right and partly wrong.

AI often generates answers from material that humans have already written down. But human scientific discoveries rarely appeared from nowhere either. New theories usually build on existing observations, instruments, mathematical language, and earlier work. What changes is the addition of a hypothesis that has not yet been tested, together with experiments that can support or reject it.

Humans do not always begin with an answer and then state it. In many important discoveries, they move through a less certain process: observe a phenomenon, identify variables, propose possible explanations, design an experiment, discover that an explanation fails, and try another one.

I keep wondering whether this process can be isolated and placed inside a controlled artificial environment.

For example, we could put an AI system in a world without Newtonian mechanics. We could start before free-fall motion had been systematically studied, provide only the observations, instruments, and mathematical resources available at that point, and let the system decide what to measure, how to measure it, what explanations to propose, and when a pattern is reliable enough to keep.

This is not about making AI pretend to be Newton

It is easy to interpret the idea as a historical role-play exercise: give a model a seventeenth-century costume, ask it to play Newton, and see whether it can recite Newton’s laws.

That experiment would tell us very little.

A modern LLM may already have encountered free fall, Galileo, Newton, acceleration, and universal gravitation in its training data. Even if the formula is absent from the prompt, the model’s parameters may still contain the concepts. Producing a narrative that moves from an experiment to a law does not mean the model actually went through the discovery.

The more precise goal is to construct a knowledge environment with no supplied answer but with the ability to run experiments, and then test whether the AI can move from observations to a law that makes testable predictions.

The important question is not which famous historical sentence it eventually produces. It is whether the path it took can be inspected.

Reconstructing the conditions of discovery

A serious experiment would need to control several kinds of conditions.

A knowledge boundary

The system should have a clear boundary around what information is available and what is not. Removing modern physics textbooks from the context is not enough. Modern concepts can leak through model parameters, vocabulary, tool descriptions, prompts, and evaluators.

Observations and instruments

The model should receive raw observations rather than conclusions. It might measure how long an object takes to travel a distance at different ramp angles and repeat the experiment, but it should not be handed the sentence “displacement is proportional to the square of time.”

The ability to design experiments

The model should not only analyze data produced by someone else. It should decide what to do next. It might change the ramp angle, distance, object mass, or measurement precision. It might repeat an experiment to find out where the error comes from.

Hypotheses and predictions

A proposed explanation is not enough. The model must turn it into a quantitative prediction and test it under conditions it has not seen before.

Records and falsification

Every experiment, failed hypothesis, and theoretical revision should be preserved. The system should not display only the final formula. It should leave behind a discovery process that another person can inspect.

Together, these conditions look more like scientific work than a question-answering task.

From falling objects to a law of motion

There is also a historical correction to make here. The systematic study of free fall is mainly associated with Galileo. Newtonian mechanics was built on work by Galileo, Kepler, Descartes, and others.

So instead of saying “let AI discover Newtonian mechanics directly from free fall,” it is more accurate to describe a knowledge chain that gradually expands.

First, the AI would need to turn motion into measurable quantities: distance, time, velocity, and differences between experimental conditions.

Next, it might compare competing mathematical relationships. Is the distance proportional to time? Is it proportional to the square of time? Does the relationship survive when the ramp angle changes?

Then it would need to test its hypothesis with an experiment that has not already been used. If a hypothesis fits the existing data but fails at a new angle or distance, the model should revise it rather than continue explaining the old observations.

Finally, it would have to identify the scope of the law. Real experiments contain friction, air resistance, and measurement error. A formula that works under ideal conditions should not automatically be presented as a universal law of motion.

The value of this setup is not that we do not know the final answer. We do. The value is that we can inspect whether the AI knew why it chose its next experiment and whether new evidence actually changed its judgment.

Hiding the textbook is not enough

This is the hardest part of the idea.

If we use a model that has already read modern science, the historical environment may be historical only on the surface. The model may never see Newton’s law in the prompt, yet still know that the words “free fall” and “acceleration” usually point toward a familiar answer.

A rigorous experiment could use at least three approaches.

The first is a knowledge-limited model, trained or filtered around the literature available before a chosen historical period. This is closer to historical reconstruction, but it is difficult to remove every trace of modern knowledge.

The second is to create randomly generated scientific worlds. Each world would have different hidden laws, unfamiliar variable names, and noisy observations. The model could not solve the task by recalling a famous formula.

The third combines both approaches: use historical environments as an interpretable curriculum, and use random worlds to test whether discovery behavior actually transfers.

The third approach is more convincing to me. Historical cases help us understand the story of discovery, while random worlds provide a stronger test.

What counts as rediscovery?

If the model outputs s ∝ t², that is evidence that it may have found a correct relationship. It is not, by itself, evidence that the model performed a scientific discovery.

We should distinguish at least the following cases:

• Answer reproduction: the model directly outputs a known formula from its training data;
• Curve fitting: the model finds an expression that explains the current data but does not predict new conditions;
• Process-level discovery: the model chooses experiments, proposes competing hypotheses, uses new data to distinguish them, and revises its theory;
• Transferable discovery: the model can discover the corresponding relationship under new units, noise levels, parameters, and simulated worlds;
• Human-new science: the result is new not only to the model but also to human knowledge, and it is validated in the real world.

These levels should not be collapsed into one claim.

A discovery that is new to the model can still be a valuable research result because it shows more than retrieval. But it should not be advertised as “AI invented a new natural science.”

Train the discovery process, not the final answer

If we want to train scientific discovery behavior, the data should not consist only of question–formula pairs. It should contain complete exploration trajectories:

observe
→ represent variables and uncertainty
→ propose a hypothesis
→ design an experiment that distinguishes hypotheses
→ collect new data
→ compare prediction with observation
→ revise or abandon the hypothesis
→ predict under new conditions

The amount of guidance should also decrease over time.

At first, we might remind the AI to list variables, controlled conditions, and possible sources of error. Later, we might only ask it to check whether its prediction has been independently tested. Eventually, we should give it a goal, tools, and a limited experiment budget without telling it which relationship to search for.

If the teacher says, “Consider whether displacement is related to the square of time,” that is a hint, not a discovery. A meaningful evaluation should compare different levels of scaffolding: no hint, a method hint, a specific directional hint, and a trained model operating in a completely unfamiliar world.

At the system level, I would separate the responsibilities:

• the LLM proposes hypotheses, explains results, and chooses the next experiment;
• the simulator returns real or noisy observations;
• a symbolic-regression tool may help search expressions, but it should not decide what scientific question the model is asking;
• the Harness controls knowledge isolation, tool permissions, experiment budgets, and result validation;
• SFT, LoRA, or RL can gradually teach better experimental design and falsification habits.

LoRA can change the model’s tendency to explore, but it cannot replace the experimental environment. A Harness can ensure that an experiment really happened, but it cannot invent a good hypothesis from nothing.

A practical MVP

The first version does not need to recreate seventeenth-century Europe, and it does not need a new large model on day one.

We could build a Scientific Discovery Gym: a two-dimensional physics environment with hidden laws.

The model could call an experiment tool, set a ramp angle, distance, friction coefficient, and release condition, and receive a time series with measurement noise, including time, position, and velocity. It would decide how to design experiments, record results, and when to propose a mathematical hypothesis.

The evaluation should measure more than whether the model writes the right formula:

• Does it control variables?
• Does it repeat experiments and handle noise?
• Does it propose competing hypotheses that can actually be distinguished?
• Does it predict new parameter settings accurately?
• Does it accept counterexamples?
• Can it transfer a law across units and randomly generated worlds?
• Does it independently reach the same or an equivalent law across multiple runs?

If every simulator follows Earth-like physics, the model may still be recalling physics. Multiple randomly generated worlds are needed to test whether it has learned to move from evidence to laws.

From free fall to Newtonian mechanics may require a curriculum

Free fall alone is not enough to derive the whole of Newtonian mechanics. A broader theory also requires inertia, mass, force, projectile motion, astronomical observations, and planetary motion.

The curriculum could therefore be divided into stages:

1. identify uniform and accelerated motion from observations; 2. study ramps and free fall; 3. analyze the horizontal and vertical components of projectile motion; 4. examine relationships between mass, external force, and changes in motion; 5. place terrestrial experiments and astronomical observations in one framework; 6. compare the predictive range and complexity of competing theories.

This curriculum should not be mistaken for a script that guarantees success. If every step is already written down, the AI is completing textbook exercises. The real research question is whether, after historical hints are reduced, it can decide what to observe next.

The connection to my earlier clarification research

My earlier LLM clarification experiment studied a smaller-looking question: when information is incomplete, should the model ask, act directly, or repair the result after failure?

The scientific-discovery environment asks a broader version of the same question. When should the AI keep observing? When should it design an experiment? When should it ask a human for missing information? When does it have enough evidence to formulate a theory?

A scientific-discovery agent may therefore need an adaptive policy rather than a fixed workflow. It should not always ask first, and it should not always act immediately. It needs to estimate the value of the next piece of information and the cost of obtaining it.

The earlier clarification experiment, however, only measured the cost and quality of interaction protocols on a limited set of text tasks. It did not show that current LLMs can discover scientific laws. Those claims must remain separate.

The conclusion

I think this direction is feasible, and it is closer to the real question than simply giving AI more examples of innovation.

The goal should not be to put a Newton costume on an AI system and see whether it can repeat Newton’s answer. The goal should be to construct an environment with no supplied answer, where the system can run experiments and test its predictions.

What we need to observe is:

Can an AI start from phenomena, propose hypotheses, design experiments that distinguish them, accept failure, and transfer a law to conditions it has not seen before?

If it only repeats formulas from familiar physics, we have retrieval and recombination. If it can discover hidden regularities in random worlds through experiments, it has at least learned part of the working method of science.

That still would not be creativity in the full human sense, and it would not mean that AI can independently discover natural science. But it could be a testable path from models that generate answers to agents that build models from evidence.