The Eurasia Review opinion piece challenges a growing assumption in AI-driven scientific research: that if a model is powerful enough, it can always extract a precise answer from experimental data. The author argues that this is fundamentally wrong. Every physical measurement contains a finite amount of information, and once an experiment does not contain enough information to distinguish between competing possibilities, no algorithm—regardless of how advanced—can recover the missing information. AI can produce an estimate, but an estimate is not necessarily a measurement.
The distinction becomes especially important when AI systems generate highly precise-looking predictions from noisy or incomplete observations. A model can learn patterns from enormous datasets and interpolate successfully within the conditions it has seen. But if the underlying experiment cannot physically resolve two possibilities, the model cannot legitimately determine which one actually occurred. The danger is that numerical precision can create an illusion of scientific certainty: a result reported to many decimal places may still rest on fundamentally insufficient evidence.
This is particularly relevant to scientific AI because models are increasingly being used for inverse problems—working backward from observations to infer properties that cannot be directly measured. In such situations, the mathematical problem may be underdetermined: multiple physical explanations can produce essentially the same observed data. AI may select the explanation that is statistically most likely based on its training, but that does not mean the data uniquely established it. The model's confidence therefore needs to be distinguished from the information actually contained in the experiment.
The broader lesson is that AI cannot repeal the laws of information and physics. Better models can reduce noise, combine measurements, identify patterns and make useful predictions, but they cannot manufacture evidence that an experiment never captured. Scientific AI therefore needs explicit awareness of measurement limits, uncertainty and identifiability. The most trustworthy systems will not simply answer “Here is the number”; they will also be able to indicate “This is what the evidence supports—and this is what the experiment cannot determine.” That distinction could become crucial as AI moves from assisting scientific analysis toward making increasingly consequential scientific claims.
