Let me tell you about the most dangerous idea in modern statistics. It was invented by a man who believed in miracles.
Thomas Bayes was a Presbyterian minister in 18th-century England. He published exactly two works in his lifetime: one on theology, and one on probability. The probability paper was so obscure that it was only published after his death by a friend. And yet, that single paper has become the mathematical foundation of AI, drug discovery, spam filters, and even the way Google ranks search results.
Here’s the irony that nobody talks about: The only person who told statisticians that their prior beliefs mattered was a man who believed in God.
Think about that. In a field that prides itself on objectivity, on data-driven decisions, on the cold, hard numbers—the foundational insight is that you can’t escape your starting assumptions. Every Bayesian analysis begins with a “prior”—a subjective probability that represents what you already believe. And then you update that belief with new data. But the prior is always there, like a ghost in the machine.
I’ve spent years in data science, and I’ve watched teams argue over priors like theologians arguing over scripture. They’d pick a “neutral” prior and pretend it wasn’t a choice. But neutrality is a prior too. It’s just a lazy one.
Bayes didn’t invent a tool. He invented a confession: that all knowledge starts with faith.
This is the twist that makes statisticians uncomfortable. The frequentist school—the other major statistical philosophy—tries to avoid priors altogether. They say, “Let the data speak for itself.” But that’s a lie. Every experimental design, every sampling method, every hypothesis test is built on assumptions. The question isn’t whether you have a prior. It’s whether you’re honest about it.
I saw this firsthand when I worked on a machine learning project for a healthcare company. The team spent weeks tuning a Bayesian model to predict patient outcomes. But the prior we used was based on data from one hospital—a wealthy, urban hospital. When we deployed the model in a rural clinic, it failed catastrophically. Not because the math was wrong. Because the prior was wrong. The Reverend would have understood.
Bayesian inference is not just a technical tool. It’s a philosophical acknowledgment that all knowledge starts with prior assumptions—even in fields that claim to be value-free. Every time you use a Bayesian filter, you’re trusting a reverend’s theology.
This matters because AI is now making life-or-death decisions. Self-driving cars, medical diagnoses, criminal justice algorithms—all of them rely on Bayesian methods. And the priors embedded in those systems carry the biases of their creators. If you don’t examine the prior, you’re not doing science. You’re doing ritual.
So here’s my take: embrace the irony. Statistics was never objective. It was always a conversation between belief and evidence. Bayes gave us the language for that conversation. The least we can do is stop pretending it’s pure math.
Don’t let the data fool you. The question is always: what did you believe before you saw it?
FAQ
Q: Doesn't Bayesian statistics just use prior probabilities that can be derived from data?
A: In theory, yes. But in practice, priors are always chosen by a human. A 'non-informative' prior is still a choice. The key is that Bayes forces you to acknowledge that choice, which is why many statisticians prefer frequentist methods that hide their assumptions.
Q: How does this affect me as a data scientist or AI engineer?
A: It means you should always question the priors in your models. If you're using a Bayesian neural network, ask: what prior assumption did I make about the weights? And if you're using a frequentist method, ask: what design assumptions did I implicitly make? The answer is never 'none.'
Q: Isn't this just a philosophical nitpick? The math works either way.
A: The math works under certain assumptions. But the history of scientific failures—from p-hacking to the replication crisis—shows that ignoring assumptions leads to overconfidence. Bayes' insight is that being honest about priors actually makes your conclusions more robust, not less.