In May 1919, Arthur Eddington photographed stars near the edge of the eclipsed Sun. Einstein's general relativity predicted their light would bend by about 1.75 arcseconds. Newtonian reasoning suggested roughly half that. The measurements were messy and the instruments imperfect, yet many physicists came away more confident in Einstein. But how much more confident should they have been?

That question sits at the heart of Bayesian reasoning, a framework built on a theorem by the minister Thomas Bayes, published after his death in 1763. Its core idea is simple: belief comes in degrees, and evidence should adjust those degrees in a disciplined way. Over the past century, this idea has reshaped how philosophers understand theory testing. It has also offered fresh answers to some of science's oldest methodological puzzles.

Classical accounts of scientific method often treat each experiment as if it stood alone. Bayesian reasoning rejects this. Before any new evidence arrives, a scientist already holds some degree of confidence in a hypothesis. This is the prior probability, a numerical expression of what is already known, from established theory to past experimental track records. A prior is not a prejudice. It is a summary.

Consider the 2011 OPERA experiment, which appeared to show neutrinos travelling faster than light. The data were carefully gathered, yet most physicists remained skeptical. Their prior confidence in special relativity, supported by a century of confirmations, was enormous. Months later, the anomaly was traced to a loose fibre-optic cable. The skepticism was not stubbornness. It was background knowledge doing its proper work.

What Bayesianism contributes is formality. Scientists have always weighed new results against what they already believed, but often informally and inconsistently. Bayes' theorem makes this weighing explicit. It also explains a familiar slogan: extraordinary claims require extraordinary evidence. A hypothesis with a low prior needs evidence that is very unlikely under its rivals before it can earn serious confidence.

Takeaway

The question is never whether to use background knowledge, since every scientist does. The real question is whether to use it openly, where others can examine and challenge it.

Bayes' theorem describes how a prior becomes a posterior probability once evidence is in hand. The key ingredient is the likelihood ratio. This compares how probable the evidence is if the hypothesis is true with how probable it is if the hypothesis is false. Evidence that a theory and its rivals predict equally barely moves the needle. Evidence that only one theory predicts can shift confidence dramatically.

The history of continental drift illustrates this gradual process. When Alfred Wegener proposed it in 1912, matching coastlines and fossils gave it modest support, but no plausible mechanism existed. Over decades, further evidence accumulated. Rock formations were found to align across oceans. Paleomagnetic records followed. Finally, in the 1960s, researchers found symmetric magnetic stripes created by seafloor spreading. Each finding raised confidence, until plate tectonics became the consensus.

This picture captures something Karl Popper emphasised: risky predictions matter most. A theory that forecasts an unexpected result gains more when that result appears, because the result was improbable otherwise. Where Bayesians and Popper part ways is on confirmation itself. Popper held that theories are never made more probable, only tested. Bayesians argue that confidence can rise legitimately, one observation at a time.

Takeaway

Confidence in a scientific theory is rarely won in a single decisive moment. It accumulates, and the evidence that counts most is the evidence that would have been surprising if the theory were wrong.

The most common objection to Bayesianism concerns its starting point. If priors reflect each scientist's judgment, two equally competent researchers might assign very different initial probabilities to the same hypothesis. Thinkers such as Bruno de Finetti and Leonard Savage embraced this openly. They treated probability as a measure of personal belief. Critics worried this made science a matter of opinion.

The Bayesian reply rests on a mathematical result often called the washing out of priors. Under suitable conditions, people who begin with different priors but agree on how likely the evidence is under each hypothesis will see their posteriors converge as evidence accumulates. A skeptic and an enthusiast studying the same data will, given enough of it, end up close together. Shared evidence comes to dominate.

The conditions matter. Convergence fails if someone assigns a hypothesis a prior of exactly zero, since no evidence can then revive it. It also requires agreement on likelihoods and enough data. Still, for a scientific realist, the result is significant. Minds starting from different places can be driven toward the same conclusions. That suggests the evidence is tracking something independent of us.

Takeaway

Objectivity need not mean starting without opinions. It can mean committing to a process where honest evidence eventually overrules where anyone began, provided no one closes the door completely.

Bayesian reasoning does not settle every debate in philosophy of science. Questions remain about how to assign priors fairly. Evidence that was already known before a theory was proposed is also hard to handle. But Bayesianism offers a precise language for what careful scientists have long done by intuition.

By making background knowledge, evidence, and disagreement explicit, Bayesianism shows that scientific confidence is neither arbitrary nor absolute. It is a measured response to the world, and it sharpens as the world keeps answering back.