How do we separate science from everything else? It sounds simple, but this question, known as the demarcation problem, is a central puzzle in the philosophy of science. It’s not just about sorting astronomy from astrology. The line we draw determines which disciplines get funding, whose expert testimony is trusted in a courtroom, and which fields we look to for reliable knowledge about the world.

This isn't an abstract debate. For a rapidly evolving field like Artificial Intelligence, the question is urgent. Is building a large language model a scientific act? Is the study of AI a science like biology or physics, or is it a form of engineering, or something else entirely? To be considered a credible science, a field needs clear criteria for what counts as a valid contribution. Simple observation isn't enough.

In the early 20th century, a group of philosophers and scientists tried to solve the demarcation problem once and for all. Known as the logical positivists, they were deeply impressed by the progress in physics and wanted to put all knowledge on a similarly firm footing. Their main tool was a principle called verificationism.

The idea was straightforward: a statement is only scientifically meaningful if it can be verified through empirical observation. In other words, if you can't test it with an experiment or observation, it's not science. It might be poetry, it might be metaphysics, but it doesn't belong in the realm of scientific knowledge. This group, gathering in Austria, became known as the Vienna Circle and their influence was immense.

This principle has a strong intuitive appeal. The claim "copper conducts electricity" can be easily tested. The claim "stealing is morally wrong" cannot be tested in the same way. The first is a scientific statement; the second belongs to ethics. It seemed like a powerful way to weed out untestable speculation and what we now call pseudoscience from legitimate inquiry.

However, verificationism soon ran into fatal problems. One major issue is that it rules out universal laws. The statement "all electrons have a negative charge" is a fundamental law of physics, but it's impossible to verify completely. You would have to test every single electron in the universe, which is impossible. You can only confirm it for the electrons you've observed.

Even more damning, the principle of verification itself cannot be empirically verified. You can't design an experiment to prove that "a statement is only meaningful if it can be empirically verified." By its own logic, verificationism is meaningless.

The struggles of logical positivism have direct implications for AI. If we tried to apply a strict verificationist standard to AI, many of its core concepts would fail. Consider the claim that a neural network has 'learned' to recognise cats in images. We can verify its performance, sure. We can show it a million pictures and see that it correctly identifies cats 99% of the time. But have we verified that it understands what a cat is, in the way a human does? No.

That deeper claim is not directly testable. This is why the demarcation problem remains so important. For AI to establish itself as a robust science, it must be clear about which of its claims are empirically testable and which are interpretations. We test performance and behaviour, not internal, subjective states like 'understanding'.

While verificationism as a strict rule failed, its spirit lives on. It forced science to focus on testability and empirical evidence. In AI research, this legacy is seen in the emphasis on benchmarks, datasets, and measurable performance metrics. Researchers don't just claim their model is 'better'; they prove it by demonstrating superior performance on standardised tasks.

This commitment to rigorous, empirical validation is what separates scientific AI research from speculative fiction about artificial minds. The line between science and non-science is not always bright, but the drive to make claims that can be tested against evidence remains a core scientific value.

Time to check your understanding of these core ideas.