When the Evidence Says You’re Wrong: How STEM Learners Should Revise a Model

By Andrew B. Raupp, Founder & Executive Director, STEM.org Educational Research™

Short answer: strong STEM thinking is not the ability to defend a first answer forever. It is the ability to decide when evidence is strong enough to change a model—and when an unexpected result deserves another test before anything is overturned.

One of the most important moments in STEM is also one of the least comfortable:

The data do not behave the way you expected.

The bridge bends more than the calculation predicted. The code returns the wrong value. The plant does not respond to the treatment. The graph moves in the opposite direction. A measurement falls outside the expected range. A simulation and an experiment disagree.

At that point, a learner has a choice.

They can protect the first idea. They can throw it away too quickly. Or they can do the more difficult work of figuring out what, exactly, the contradiction means.

That third response is one of the defining habits of serious STEM work.

An anomaly is information, not a verdict

In science education research, contradictory or anomalous data have long been studied as a trigger for conceptual change. The intuitive assumption is simple: show someone evidence that conflicts with an incorrect idea and the idea should disappear.

That is not what reliably happens.

Clark Chinn and William Brewer’s work on anomalous data showed that people can respond in many ways when evidence conflicts with an existing belief. They may reject the data, exclude it as irrelevant, question the methods, reinterpret the result, make a small adjustment that preserves the core model, or genuinely revise the model itself.

That matters because the educational goal is not to train students to surrender whenever they see one surprising result. A single anomaly can come from a bad sensor, a transcription error, an uncontrolled variable, a coding bug, an inappropriate model, random variation—or a genuinely mistaken theory.

Good skepticism works in both directions.

Do not dismiss evidence merely because it is inconvenient. But do not abandon a well-supported model because one questionable observation looks dramatic.

The real question is: what has to change?

A productive response to conflicting evidence begins by localizing the disagreement.

Suppose an engineering team predicts that a small structure will hold 20 kilograms, but it fails at 12. The useful question is not simply, “Was the prediction wrong?”

More useful questions are:

• Was the load measured correctly?
• Did the material match the assumed properties?
• Was the geometry built to the intended dimensions?
• Was an important force omitted from the model?
• Did the model use the wrong boundary condition?
• Does the failure repeat?

Each question points to a different kind of revision.

Sometimes the data need correction. Sometimes the procedure does. Sometimes an assumption does. Sometimes the underlying model has to change.

Learning to distinguish those possibilities is much more valuable than learning the slogan “follow the evidence.” Evidence has to be interpreted, tested and situated.

Why misconceptions can survive good explanations

Research on conceptual change helps explain why simply telling students the correct answer often has limited power.

Prior knowledge is not an empty slot waiting to be overwritten. It can remain active even after a correction has been presented. Panayiota Kendeou’s Knowledge Revision Components work examines the conditions under which new information can compete successfully with prior, often incorrect knowledge during learning.

This means a student can repeat the new explanation and still fall back on the old one when the context changes.

That is why model revision should involve more than exposure to the correct statement. The learner needs to notice the conflict, understand why the old explanation fails, see why the replacement explains the evidence better, and use the revised model again.

A 2024 meta-analysis by Çisem Paçacı, Uğur Üstün and Ömer Faruk Özdemir examined conceptual-change strategies in science education, including cognitive conflict, cognitive bridging and ontological category shift. The existence of these different strategies points to an important practical reality: misconceptions are not all the same, and there is no single correction technique that fits every case.

A Model Revision Loop for STEM learning

Students do not need a philosophy-of-science seminar every time a result looks odd. They need a repeatable habit for deciding what to inspect next.

What this looks like across STEM

In mathematics, a student may discover that a procedure works for several familiar examples but breaks under a negative number, a zero denominator or a different representation. The exception is not a nuisance. It reveals the boundary of the rule.

In science, an unexpected measurement should lead first to questions about repeatability, controls and instrumentation before it becomes a sweeping conclusion.

In engineering, failure is often a model update in physical form. A prototype shows which assumption about material, load, tolerance or user behavior did not survive contact with reality.

In coding, a failing test case is evidence about a mental model of the program. Debugging improves when a student predicts where the failure must originate instead of making random changes until the error disappears.

In data analysis, a surprising pattern should trigger checks for denominator changes, selection effects, missing values, scale choices and data provenance before a narrative hardens around it.

Across all of these fields, the habit is the same: treat contradiction as a request for a better explanation.

AI makes model revision more important, not less

Generative AI can offer an explanation almost instantly. That is useful, but it can create a subtle problem: the learner may receive a revised answer before forming a revised model.

If an AI system says, “Your reasoning is wrong because…,” the student still needs to decide whether the explanation fits the evidence. Otherwise correction becomes another form of authority dependence.

This connects directly to the broader question of human agency in responsible AI use. A tool can accelerate inspection, propose alternative explanations or generate test cases. It should not eliminate the learner’s responsibility to judge which explanation survives.

The same principle underlies estimating before trusting an automated answer and preserving useful difficulty in STEM learning. The goal is not to protect learners from tools. It is to keep the learner intellectually present while using them.

Changing your mind is not intellectual weakness

There is a cultural temptation to treat consistency as a virtue even when the evidence changes.

STEM should teach the opposite lesson.

A model is valuable because it survives serious attempts to test it—not because someone became attached to it first.

The mature STEM learner is neither stubborn nor gullible. They do not cling to a theory because it is familiar, and they do not abandon it because a single surprising result appears on a screen.

They ask what the evidence can actually support.

Then they revise accordingly.

That is not merely a classroom habit. It is one of the foundations of sound judgment.

---

Research referenced

• Chinn, C. A. & Brewer, W. F. (1993). The role of anomalous data in knowledge acquisition: A theoretical framework and implications for science instruction. Review of Educational Research, 63(1), 1–49.
• Chinn, C. A. & Brewer, W. F. (1998). An empirical test of a taxonomy of responses to anomalous data in science. Journal of Research in Science Teaching, 35(6), 623–654.
• Paçacı, Ç., Üstün, U. & Özdemir, Ö. F. (2024). Effectiveness of conceptual change strategies in science education: A meta-analysis. Journal of Research in Science Teaching.
• Kendeou, P. (2024). A theory of knowledge revision: The development of the KReC framework. Educational Psychology Review.