AI in de wetenschap

Scientific usefulness depends on reproducibility, uncertainty, data provenance, and whether the method answers the stated question. A prediction is not automatically a discovery or a causal explanation.

Key takeaways

• State the scientific question and disconfirming evidence.
• Preserve provenance and reproducibility.
• Separate hypotheses and predictions from validated findings.

Frame the scientific question before selecting an algorithm. Decide what is measured, what is inferred, and what observation would disconfirm the claim. Keep training, validation, and test data separate, especially when measurements from the same subject, instrument, or experiment are correlated. Record preprocessing, model versions, random seeds where relevant, and evaluation material. Check whether missing data or selection effects change the conclusion. A model can reproduce a known pattern while failing on a new instrument, population, or experimental condition. Use uncertainty honestly. Calibration, confidence intervals, prediction intervals, and sensitivity analyses answer different questions. A generated hypothesis can guide follow-up work, but it is not evidence until an appropriate experiment or independent analysis supports it. Preserve the path from source data to figure, table, or manuscript. Review authorship, citations, and generated text carefully, and avoid claiming that an automated result was independently replicated when it was not.

1. Imagine a model trained on data from instrument A and tested on a random split of the same instrument’s readings.
2. A second evaluation uses later readings from instrument B and shows a large error increase.
3. Report both results and investigate calibration or measurement differences before claiming general scientific performance.

The constructed example illustrates why random splits can overstate scientific generalization.