An AI valuation model that undervalues injury claims is ordinarily treated as negligence — a flawed process, badly configured — but it crosses into gross negligence when the insurer knows the model pays too little and keeps using it anyway. The dividing line is knowledge, not error: gross negligence is defined as the absence of even slight care, or "a conscious, voluntary act or omission in reckless disregard of a legal duty and of the consequences to another party," as summarized in the standard definition.
That distinction decides real money. Ordinary negligence and breach of contract get a claimant the benefits they were owed; conscious disregard opens the door to bad-faith liability and punitive damages. The litigation and regulatory record around insurance valuation software shows exactly which facts push a case over that line.
Table of Contents
- What "undervaluing by design" looks like in the record
- The facts that convert negligence into conscious disregard
- Why the contract claim usually survives before the tort claim does
- Discovery is where the knowledge question is won or lost
- New statutes are turning the question from judgment into violation
- What to do if you think a model priced your claim
- Frequently Asked Questions
What "undervaluing by design" looks like in the record
The most concrete example is Colossus, a bodily-injury valuation program used by major auto insurers to put a dollar range on a claimant's injury. Its vendor's sales literature told insurers the program "will immediately reduce the size of bodily injury claims by up to 20 percent," a claim documented by Michigan Auto Law. That sentence matters legally because it describes reduction, not accuracy, as the product's purpose.
Regulators acted on the conduct around it rather than the math inside it. Allstate entered a Multi-State Market Conduct Regulatory Agreement with insurance commissioners in 48 states, paying a $10 million fine and agreeing to disclose Colossus use to claimants, apply uniform criteria, audit adherence, allow adjusters to settle above the software's number, and stop incentivizing adjusters to settle at it. Read the last two terms closely. They target the part that looks like conscious disregard: an adjuster who is paid to hit the model's figure has no room to correct it when it is wrong.
The facts that convert negligence into conscious disregard
The clearest modern illustration is the health-insurance side. Plaintiffs in *Estate of Gene B. Lokken v. UnitedHealth Group* allege the insurer's nH Predict algorithm carried a known error rate of roughly 90% and was used in place of physicians to cut post-acute care regardless of the treating doctor's recommendation, according to Georgetown's Health Care Litigation Tracker.
A known error rate is the hinge. An insurer that misconfigures a model and does not realize it has a process failure. An insurer holding internal figures showing the model gets the answer wrong most of the time, and running it anyway, is making a choice about consequences to another party. For a personal injury claimant, the analogous facts to look for are these:.
- Internal accuracy testing, audits, or actuarial memos showing the model pays below documented values
- Adjuster incentives, scorecards, or authority limits tied to the model's number
- Instructions discouraging or forbidding settlement above the model output
- Vendor marketing that promises savings rather than accuracy
- Continued use after regulators, courts, or internal counsel flagged the problem
Why the contract claim usually survives before the tort claim does
Courts have been readier to let a claimant proceed on the policy's own promises than on a novel tort theory. On February 13, 2025, the U.S. District Court for the District of Minnesota refused to dismiss the breach-of-contract and breach-of-implied-covenant-of-good-faith claims in Lokken, reasoning that the live question is whether the insurer complied with its own coverage documents, which promised decisions by "clinical services staff" and "physicians," while allegedly using an algorithm instead. The reasoning appears in the court's opinion.
The practical lesson transfers directly to injury claims. Start with what the insurer said it would do — in the policy, the claims handling literature, the regulatory filings — and compare it to what the model actually did. A promise of individualized evaluation, sitting next to a program that assigns a value from a dataset, is a documentary mismatch rather than an argument about software. Gross negligence then does separate work. Punitive damages in bad-faith actions require a showing of conscious disregard of others' rights, or conduct so reckless as to amount to it, and courts generally hold that negligence plus something more — gross or reckless negligence — can establish bad faith, per an analysis by the Connecticut General Assembly's Office of Legislative Research.
Discovery is where the knowledge question is won or lost
Conscious disregard lives in the insurer's files, so the fight is about access. A magistrate judge in the Lokken case later compelled UnitedHealth to produce documents on nH Predict's development and use, related government investigations, and the company's oversight of its AI systems, as reported by Hanover Hangley Law. That ruling is the template. Development records show what the model was built to do; oversight records show who was watching; government investigation files show whether the insurer had already been told.
Together they are the evidence of scienter — knowledge and intent — that a gross negligence theory needs. Be realistic about the limits. These are pleading-stage and discovery-stage rulings, not verdicts, and none of them holds that using a valuation model is itself unlawful. They establish something narrower and still useful: that allegations about a known-bad model survive dismissal and that the insurer's internal knowledge is discoverable.
New statutes are turning the question from judgment into violation
Several states have removed the discretion insurers previously had. California's SB 1120, the Physicians Make Decisions Act, was signed on September 30, 2024 and took effect January 1, 2025. It bars AI from supplanting provider decision-making on medical necessity, requires a licensed clinician competent in the specific clinical issues to make any denial, delay or modification, and forbids the tool from basing its analysis solely on a group dataset — terms set out by the office of Senator Josh Becker. Other states have followed with the same structure.
Arizona's HB 2175 requires an independent provider review before a denial and prohibits relying solely on any other source, and Nebraska's LB 77 bars AI output as the sole basis for denying, delaying or modifying care while requiring public disclosure of AI use in utilization review, as tracked by Holland & Knight. Once such a statute applies, letting the model decide is not a defensible judgment call — it is a rule violation, and that recharacterizes the conduct. A company that keeps a prohibited process running after the effective date is, on its face, disregarding a legal duty. These particular statutes govern medical necessity and utilization review rather than auto bodily-injury valuation, so check whether your claim type is covered before relying on one.
What to do if you think a model priced your claim
The steps below are about building a record, not about winning an argument with an adjuster.
- Ask in writing whether any software or automated tool was used to value or deny your claim, and ask for its name. Allstate's multi-state agreement committed it to disclosing Colossus use to claimants, so the question is not unprecedented.
- Request your complete claim file, including the valuation output and any range the adjuster was given.
- Compare the offer against your actual documented losses — medical bills, wage records, and physician opinions — rather than against the insurer's explanation of its number.
- Ask whether the adjuster has authority to settle above the tool's figure. That single answer goes to the incentive problem regulators targeted.
- Complain to your state insurance department, which is the body that produced the Colossus conduct agreement in the first place.
Frequently Asked Questions
Is it illegal for an insurer to use software to value my injury claim?
No. Regulators addressed how the software was used — disclosure, uniform criteria, audits, and adjuster incentives — rather than banning valuation programs. Some newer state statutes do restrict AI in medical necessity decisions specifically.
Does a low offer prove gross negligence?
No. A low offer alone looks like ordinary negligence or a contract dispute. Gross negligence requires evidence the insurer knew the process underpaid and continued regardless.
What is the practical difference for my recovery?
Breach of contract generally recovers what you were owed. A sustained bad-faith theory can reach additional damages, and punitive damages require a showing of conscious disregard of others' rights or comparable recklessness.