When AI Should Say It Doesn't Know

TL;DR — To accurately forecast claims, separate reading from reasoning. Use generative AI to structure thousands of pages of case files, then pass that data to mathematical models that calculate calibrated settlement ranges and escalation probabilities based on historical outcomes.

Hand a complex traumatic brain injury file to a standard artificial intelligence tool, and it will often spit out a single settlement value. Let us say the screen blinks and returns $1,450,000. This is the exact moment the technology fails. A point estimate in litigation is a mathematical fiction. A single number implies absolute certainty in a system defined by chaos. Third-party litigation funding, venue volatility, and social inflation guarantee that no single case has a deterministic outcome. When an algorithm provides a specific dollar figure without a margin of error, it is not predicting. It is guessing.

The problem stems from a fundamental misunderstanding of what large language models actually do. Language models generate text based on statistical proximity. They are next-token predictors built to produce plausible sentences, not to calculate risk. If you force a generative model to evaluate a legal claim, it will generate a number that looks like a realistic settlement because that number fits the grammatical pattern of a legal summary. Plausibility is the exact opposite of probability. A language model cannot quantify its own uncertainty. It simply writes a confident sentence containing a hallucinated number.

The Geometry of Doubt

True prediction requires strict calibration. Calibration means the model's confidence mathematically matches reality. If an algorithm states there is a 70 percent chance a claim will escalate to trial, then out of one hundred similar claims, exactly seventy must escalate. If it predicts a settlement range of $500,000 to $800,000 with 90 percent confidence, the final outcome must fall inside that boundary nine times out of ten. This requires conformal prediction techniques and strict statistical bounds. Language models lack this architecture entirely. They have no internal mechanism to measure their own correctness.

Litigation is a distribution, not a point. A single case has a cluster of probable outcomes based on specific variables. Conformal prediction allows us to map this distribution and draw a precise boundary around the likely results. We calculate a region of certainty. If the underlying data is noisy, sparse, or contradictory, the mathematical region expands. The model expresses its doubt by widening the settlement range. This is honest error reporting. The width of the range is just as informative as the numbers themselves.

The demand for this mathematical rigor is accelerating. Claims organizations face an environment warped by social inflation and rising nuclear verdicts. Plaintiff firms utilize sophisticated data operations to maximize settlement values and push the boundaries of historical norms. In this environment, an uncalibrated point estimate is a massive liability. You cannot set a reserve based on a hallucinated average. You need a quantifiable boundary of risk that accounts for the extreme volatility of the modern judicial docket.

Reading Versus Reasoning

To achieve this accuracy, we must strictly separate reading from reasoning. A modern case file contains thousands of pages of pleadings, medical records, demands, and correspondence. No human professional can parse this volume of unstructured data instantaneously. We use generative AI to read these documents. The neural network extracts the facts, identifies the injuries, maps the involved parties, and structures the timeline. It does the heavy extraction work of unstructured data processing. Then, we strictly forbid it from making a prediction.

We employ a neural-symbolic architecture. The structured output from the reading phase flows into separate geometric machine-learning models. These algorithms are trained strictly on large numbers of resolved cases with known outcomes. They map the extracted facts against historical data to calculate a probability distribution. The output is a calibrated settlement range, an escalation probability, and a reserve delta compared to the current baseline. We replace grammar with geometry.

The predictive engine relies on mathematical distance between comparable resolved cases, entirely divorced from the language model that read the initial file. End-to-end deep learning fails in insurance because it blends the reading and the reasoning into a single black box. By separating them, we ensure that the final calculation is grounded in actual historical settlements, not linguistic patterns.

Honest Error Reporting

A calibrated range is useless if it remains a black box. A forecasting model must be able to explain its uncertainty mathematically. If a settlement range is unusually wide, the claims executive needs to know exactly why. Our mathematical model isolates the specific drivers pushing the upper boundary. It points to a particular comorbid condition in the medicals, a specific plaintiff attorney tactic, or a pattern of nuclear verdicts in the assigned venue. Every single factor is traceable directly back to the source documents.

The uncertainty itself becomes a strategic signal. A wide range tells the defense team exactly where they need to focus their discovery efforts to narrow the risk. If the model indicates a high escalation probability driven by an undocumented spinal injury claim, the adjuster knows to order an independent medical examination immediately. You allocate defense spend based on mathematical drivers rather than intuition or habit.

Claims organizations need to set realistic reserves on day one. When a forecasting model gets a prediction wrong and fails to report its uncertainty, the resulting reserve volatility damages the entire portfolio. A calibrated range allows executives to negotiate from a position of data rather than gut instinct. You know the exact probability of an adverse outcome before you enter mediation, and you have the comparable resolved cases in hand to prove it.

We must demand honest error reporting from our systems. A forecasting engine must know its limits and quantify its doubt. If the data is sparse, the algorithm must return a wider range and clearly state its lack of confidence. An algorithm that never admits uncertainty is simply lying to you in code.