The insurance industry has always been in the business of prediction. For centuries, that meant actuaries hunched over mortality tables, adjusting premiums based on age, occupation, and the occasional medical exam. Today, many of those calculations happen in milliseconds, executed by machine learning models trained on datasets so vast that no human could process them in a lifetime.
This is not speculative futurism. It is the present reality at most major insurers, and it represents one of the most consequential deployments of artificial intelligence in any industry — consequential because insurance touches nearly every economic transaction, from the mortgage on a house to the liability policy on a rideshare driver.
The new underwriting stack
Traditional underwriting relied on categorical thinking: smoker or non-smoker, urban or rural, professional class or manual labor. Machine learning models operate differently. They identify correlations across thousands of variables simultaneously, many of which would never occur to a human underwriter.
An applicant's credit behavior, purchasing patterns, social media activity, and even the way they fill out online forms can feed into risk scores. Some insurers use telematics data from vehicles to assess driving habits in real time. Others analyze satellite imagery to evaluate property risk without sending an inspector. The models do not explain their reasoning in human terms; they simply output a probability and a price.
For insurers, the efficiency gains are substantial. What once required days of human review can now happen at the point of sale. Fraud detection has improved. Pricing has become more granular, which in theory means lower premiums for lower-risk customers.
The fairness problem
But granular pricing is a double-edged sword. When algorithms can identify risk at the individual level, the traditional insurance mechanism — pooling risk across a population — begins to erode. Those deemed higher risk pay more; those deemed lower risk pay less. The healthy subsidize the sick less than they once did.
Regulators in several jurisdictions have grown concerned. The core issue is that machine learning models can reproduce and even amplify historical biases without explicitly using protected characteristics like race or gender. A model might never see an applicant's ethnicity, but if it weighs zip code, education level, and purchasing behavior, it may arrive at outcomes that correlate uncomfortably with demographic lines.
Some insurers have responded by auditing their models for disparate impact. Others argue that actuarial fairness — charging each person according to their individual risk — is the only defensible standard. The debate is far from settled, and it reveals a tension at the heart of algorithmic decision-making: statistical accuracy and social equity do not always point in the same direction.
What policyholders should understand
Most consumers have little visibility into how their premiums are calculated. The opacity is partly technical — explaining a gradient-boosted decision tree to a layperson is genuinely difficult — and partly strategic. Insurers are not eager to reveal the precise factors that influence pricing, lest applicants game the system.
This asymmetry matters. A person denied coverage or quoted an unexpectedly high premium may have no practical way to understand why, let alone contest the decision. The right to explanation, enshrined in some data protection regimes, is difficult to enforce when the explanation itself is a matrix of weighted features.
Our take
Insurance underwriting is a useful case study in how AI transforms industries without fanfare. There are no viral demos, no existential debates about sentient machines — just a quiet recalibration of who pays what for protection against misfortune. The efficiency is real, but so are the questions about accountability and fairness. As algorithms assume more decision-making authority, the burden on regulators and the public to understand them grows heavier. Most people will never read a white paper on actuarial machine learning. They will simply receive a quote and wonder why it is higher than their neighbor's.
