In an essay for Semiconductor Engineering, a writer with experience spanning structural biology, medicine, education, and chip verification argues that AI has made a long-standing question louder: how do we know something is true? The output of generative AI reads fluently and appears complete, yet the author notes a quiet friction: before any decision that matters, you still have to ask whether the result can be trusted enough to act on it. That question, once confined to laboratories and verification labs, now reaches boardrooms and classrooms.
The essay draws a direct line between disciplines that look unrelated. In science, AlphaFold's protein-structure predictions are a genuine leap, but an independent assessment found that even the most confident predictions can deviate from experimental structures by more than two angstroms in about 10% of cases. The author, a former experimental structural biologist, stresses that a model must correspond to evidence rather than merely appear plausible. In medicine, diagnostic confidence was always a chain of evidence, not an intuition, and AI-assisted reads still require a human to take responsibility for the decision. The same discipline, the author argues, governs verification in engineering: build only what the data can carry.
The piece also takes a clear stance on education: chasing AI-generated essays with better detectors misreads the problem. Detection is a policing frame; what has changed is that understanding can no longer be inferred from a finished artefact alone. The author, who spent over nine years teaching at Queen Mary University of London, frames this as an improvement challenge rather than a plagiarism problem. For the compute and hardware world, the broader lesson is that as AI accelerates the generation of models, code, and designs, verification becomes the scarce skill—and the real bottleneck to acting on what AI produces.