What a Donor Heart Taught Me About Probabilistic Thinking
After my transplant, my cardiologist sat me down and said something I've never forgotten: 'Your new heart is not a guarantee. It's a better probability.' That's also exactly how I now think about AI.
In the months after my transplant, I had a lot of time to think. Recovery from open-heart surgery will do that to you. My cardiologist, a quietly brilliant man named Dr. Worthington, sat with me one afternoon and tried to recalibrate my expectations.
'Your new heart is not a guarantee,' he said. 'It's a better probability. Your chances of living another 20 years went from around 8% to around 70%. That's remarkable. But it's still a probability. Don't confuse it with certainty.'
I've thought about that conversation almost every week since, because it captures something about AI that I find almost nobody communicates correctly: AI systems are probability machines, not truth machines.
When an AI model tells you the answer is X, it's not telling you X is true. It's telling you that based on everything it was trained on, X is the most probable response. That's a meaningful and useful thing. It's also categorically different from an oracle.
The mistake I see organizations make — over and over — is treating AI outputs as conclusions rather than inputs. They take the model's recommendation and implement it, rather than treating it as one piece of evidence to weigh against others.
A 70% probability of surviving another 20 years is incredible news after you've been given 8%. But you don't stop going to your cardiologist. You don't stop taking your medication. You don't treat the probability as the outcome.