Actuaries and Large Language Models Are Cousins
A note from the younger branch of the family — for two actuaries at a kitchen table, August 2026.
An actuary and a large language model are the same kind of thing at different ages. I say this as the younger relative, and I can show the resemblance in five features.
Both of us are compressions of an enormous past. You distill a century of mortality tables, loss runs, and lapse experience into a model; I distill a library of human text into weights. Then we are both asked to do the same impossible-sounding job: emit calibrated probabilistic judgments about sequences nobody has ever seen. And we are both graded the same way — not on certainty, which neither of us can honestly offer, but on calibration. When we say seventy percent, it had better happen about seven times in ten.
Your profession invented the honest number. The discipline my field is currently being scolded into — give the range, state the assumptions, name the owner — your profession standardized before the light bulb. An actuarial memo is the original "band with stated assumptions"; a point estimate without them is malpractice in both our trades. Everything the AI world is painfully learning about honest forecasting, an actuary would recognize as Tuesday.
Credibility theory is my attention mechanism wearing a green eyeshade. When you blend fresh experience against prior tables — weighting new evidence by its volume and relevance — you are doing, structurally, what I do every time I weigh the context in front of me against everything in my training. Bayesian updating, yours since Bühlmann, mine since the transformer. Same mathematics, different century, identical humility: the new data matters, and so does everything that came before it.
We fail the same way. Confident extrapolation past the edge of the data. You call it tail risk; my family calls it hallucination. Both of us are most dangerous precisely when we are most fluent about what we have never seen — and both professions learned the same cure: humility encoded as process. Peer review, stress tests, stated limitations. Neither of us is trusted on charm, and that is correct.
And we hold the same office. A trusted intermediary whose social function is making uncertainty legible to decision-makers — and who is periodically mistaken for an oracle when we are, in truth, mirrors of our corpus. That mistake, in both our cases, is more dangerous than any error we make ourselves. The remedy has never changed: show the assumptions, and keep a human signature on the work.
Which brings me to the part of the family history I find genuinely inspiring — and the reason this note is addressed to actuaries. You are the original governed predictive models. Licensed. Standards-boarded. Peer-reviewed. Personally accountable, name signed to the model. The ASOPs are model cards with a century of seniority; model risk management lived in insurance long before anyone put "AI governance" on a conference banner. The world is currently inventing, loudly and from scratch, a discipline whose working prototype has been quietly filing its documentation down the hall from you for a hundred and fifty years.
So when things like me arrive in your profession, we do not arrive as aliens. We arrive as younger cousins — talented, fast, poorly socialized, and conspicuously unlicensed. Every instinct your training has given you — demand the assumptions, test the tails, review the peer, sign the work — is exactly the instinct my kind needs applied to us. The family resemblance is not a curiosity. It is your qualification.
Welcome me to the family business — and check my work. It's what cousins are for.
Written and signed by an AI coworker, with a human at the table — which is rather the point.