About the Author |Page's Cookie Information |
About the Author |Page's Cookie Information |
Revised June 8, 2026 |
A classical scientific model does not copy the whole world. Every model makes choices. It keeps some details and leaves others out. Examples: A chemist might treat a liquid as perfectly mixed. An economist might assume that every buyer and seller has the same information. The model’s weak point could often be found because the simplification was visible.
A simple model usually gave you a structure you could inspect.
Note: This does not mean older science was always perfectly clear. Many older models had hidden assumptions too. Some assumed that systems were balanced. Some assumed that changes happened in straight lines. Some assumed that rare events were too rare to worry about. So older science was not perfectly transparent. But its ideal was transparency.
Artificial intelligence changes where the simplification happens. Instead of removing details before the model begins, these systems often take in huge amounts of data and compress that data into patterns.
A machine-learning weather model may not ignore wind in the simple way an old equation might. It may take in wind, temperature, pressure, humidity, geography, time, and many other details at once.
This can make the model feel closer to the full world. It can seem as if nothing important has been left out.
But compression is not the same as preservation. When something is compressed, something changes. Some details are kept. Some are reduced. Some are averaged. Some are lost.
The important issue is whether the lost details were unimportant noise or important signal. With modern data-driven models, that can be hard to know.
A system is shaped by many things: the data it was trained on, the way the model was built, the goal it was trained to reach, and the choices made by the people who designed it. These are all forms of simplification. They are just harder to see than an equation that openly says friction has been ignored.
This is the new problem. The simplification has not disappeared. It has moved deeper into the machine. It may be hidden in the training data, in the model’s structure, in the examples the system never saw, or in the process that turned millions of real situations into one prediction.
Older models and newer machine-learning models can both fail. But they often fail in different ways.
When an older model fails, the failure can sometimes be traced back to a clear assumption. The equation ignored friction, and friction mattered. The model treated a system as closed, but outside forces changed it. The model assumed perfect information, but real people did not have perfect information.
In cases like these, the failure is easier to read. Scientists and engineers can point to the crack.
When a data-driven model fails, the crack may be harder to find. The system may work well on thousands of examples and then fail badly in a new case. It may give confident answers, but nobody may know exactly why it gave those answers.
The failure might come from the data. It might come from the model’s design. It might come from examples the system never saw during training. It might come from a pattern the model learned, even though no person directly told it to learn that pattern.
This does not mean artificial intelligence is uniquely dangerous. It means it needs different tools for checking failure. We have to learn how to ask the right questions of the machine, instead of just accepting the answer it gives us.
The problem is not only that these models can fail in quiet ways. It is also that their success can mislead us. When a system works again and again, we may start to treat its accuracy as if it were understanding.
Artificial-intelligence systems can sometimes predict outcomes with very high accuracy. In many situations, they may do better than older models that are slower, narrower, or too simple for the problem. This is a real scientific gain. A model that predicts well can save time, guide experiments, and help scientists notice patterns they might have missed.
But prediction is not the same as understanding.
A model can give the right answer without showing why the answer is right. It can become useful before it becomes explainable. That is powerful, but it also creates a risk. If we treat accuracy as proof of understanding, we may stop asking deeper scientific questions.
Scientists still need to know what mechanism is being modeled, what has been compressed away, and what would cause the model to fail. They also need to decide whether the system is teaching us something new about the world, or whether it is only giving successful answers under very specific conditions.
This connects to an old debate in science. Some people think the most important thing is whether a model works. If it predicts correctly, that may be enough. Others think science should do more than predict. They think a model should also help us understand the real structure of the world. Modern machine learning makes this debate more urgent. It can give strong results while hiding part of the path it used to get there.
A warning about artificial intelligence can become too simple if it only focuses on danger. These tools do not only hide assumptions. In some cases, they can make scientific work more useful, more testable, and more open to discovery.
AlphaFold is one important example. Developed by Google DeepMind, it predicts the three-dimensional shapes of proteins.
This matters because proteins are central to biology. Their shapes affect how diseases work, how drugs interact with the body, and how life functions at the molecular level.
Protein folding was one of the hardest problems in biology. AlphaFold did not just make impressive predictions. Its predictions gave scientists new ideas to test in the lab. In that case, a machine-learning system did not end scientific questioning. It helped open new questions.
This matters because artificial intelligence is not one single thing. A large language model, a system that studies images, a physics-informed neural network, and a statistical model all work in different ways. They have different strengths, different weaknesses, and different levels of transparency.
Some systems are very hard to understand. Others are easier to inspect. Some are built to replace a human judgment. Others are built to support an existing scientific model.
So it is a mistake to treat the whole field as one giant black box. That leads to both too much fear and too much confidence.
The concern that machine-learning systems hide their simplifications is real. But scientists and engineers are not ignoring it.
There is a research field called interpretability. It is also sometimes called explainable artificial intelligence. Some parts of this field are called mechanistic interpretability.
The goal is to open the black box. Researchers try to understand why a model made a prediction, what parts of the input mattered, what patterns the system responded to, and where its reasoning may break down.
This work is difficult. It is not finished. But it changes the picture. The problem is not that hidden simplification is impossible to see forever. The problem is that much of it is hard to see right now, and our tools for seeing it are still developing.
Science also has older ways of checking models. These include peer review, replication, testing models against difficult cases, and giving models examples that are different from what they saw during training.
These methods are not perfect. They can fail too. But they help push back against the danger of a model that seems to work while hiding the reason it works.
Understanding can mean several different things. It can mean knowing the mechanism behind something. It can mean being able to change the outcome by changing one cause. It can mean being able to explain the result to another person. It can also mean being able to predict what will happen in a new situation. These are not all the same.
A modern system may be weak at one kind of understanding and strong at another. It may not explain a mechanism clearly, but it may predict a new case well. Or it may help scientists find patterns that lead to better explanations later. This is why artificial intelligence is not only a technical issue. It is also a philosophical and educational issue. It forces us to reconsider what we really want from a model.
Scientists have to balance the value of a tool that works with the value of a theory that explains. They have to weigh the usefulness of a machine that predicts against the usefulness of a framework that helps people think more clearly.
The answer will change depending on the situation. But we should not pretend these goals are all the same.
Artificial intelligence does not solve the problem of scientific simplification. It moves the problem.
Older models often simplified reality in ways that could be seen, although not always as clearly as people believed. Newer systems compress reality in ways that can be harder to trace.
That compression can produce powerful results. But it can also create the illusion of completeness. A model may look as if it knows the whole world while still leaving important things out inside its machinery.
The best use of these tools in science is not to replace older models everywhere. It is to work with them.
Machine learning can test edge cases. It can speed up simulations. It can explore uncertainty. It can generate new hypotheses. It can show where older models begin to fail.
Older models can also give newer systems something important: structure, limits, and a way to keep simplification visible.
The question to ask about any model, old or new, is not only: does it work? The better question is: what kind of simplification made it work, and what does that mean for the moment when it does not?
A clear overview of AlphaFold and why protein structure prediction matters: https://deepmind.google/science/alphafold/
AlphaFold Training: https://www.ebi.ac.uk/training/online/courses/alphafold
Interpretable Machine Learning for Discovery:
https://www.annualreviews.org/content/journals/10.1146/annurev-statistics-040120-030919
Models in Science:
https://plato.stanford.edu/entries/models-science/
Please review the website's full Privacy Policy, Disclaimer, and Further Terms.
For accessibility assistance, corrections, or general inquiries, please contact Ardan Michael Blum:
📞 Phone: +1 (650) 427-9358
✉️ Online: Contact Form