AI underwriting: what the enforcement record shows buyers in 2026
Four public actions on AI underwriting in lending and insurance. The one model a regulator tested at scale came back clean; the others turned on an input, a hard rule or a notice step. That is where a buyer should look first.
If you are buying AI underwriting, every vendor deck answers the question you were told to worry about: is the model biased. The public record says that is the wrong first question.
TIN has four case files where an AI underwriting or claims-triage system met a regulator or a court. Read together, they point somewhere less obvious than the model.
The headline figure is a settlement. Massachusetts reached a $2.5 million settlement with student lender Earnest over AI models it said could harm Black, Hispanic and non-citizen applicants [source].
What AI underwriting means
AI underwriting is the use of algorithmic or machine-learning models to decide whether to accept an applicant and on what terms, for a loan or an insurance policy.
In the four cases below it covers three jobs: deciding eligibility, setting price and terms, and picking which customers get pulled for further review.
The proof: four actions, one pattern
Each of the four is a TIN case file, checked against the public record. Each shows where the liability actually landed.
Earnest (lending, Massachusetts, 2025). The Attorney General said Earnest’s models used a “Cohort Default Rate” variable tied to an applicant’s school, and a “Knockout Rule” that automatically denied applications based on immigration status [source]. The settlement required Earnest to drop both. These are allegations resolved by settlement, not a court finding. Case file.
GEICO (auto insurance, Pennsylvania, 2026). The Attorney General said GEICO used a tool with artificial intelligence features to select a West Philadelphia policyholder for further review and underwriting, leading to cancellation without adequate notice [source]. The state’s concern was the fairness and confusion of the process, not a biased score. GEICO admitted no violation. Case file.
Apple Card (credit, New York, 2021). After a viral sex-bias allegation, DFS analysed underwriting data for approximately 400,000 New York applicants and found no unlawful discrimination [source]. It faulted customer service and transparency instead, and the bank dropped a six-month wait before applicants could appeal credit terms [source]. Case file.
Huskey v. State Farm (homeowners claims, Illinois, 2023). This one is claims handling, not underwriting, and it is the nearest insurance analogue on the record. Plaintiffs alleged State Farm’s antifraud claims tools sent Black policyholders into extra scrutiny; the court let the Fair Housing Act 3604(b) claim survive dismissal [source]. Nothing is proven yet. Case file.
Is AI underwriting risky because the model is biased?
No, not on this record. The obvious reading is that AI underwriting fails when the model learns bias. The only model here tested statistically at scale, Apple Card, came back clean [source].
What regulators acted on were things a person chose and could have checked: a schools variable, a hard immigration filter, an untested model, a cancellation notice, an appeals delay. Massachusetts also alleged Earnest failed to test its models for disparate impact at all [source].
That is the argument of this post, and it can be wrong. Four cases is a small record, and Huskey is still live. But none of the four turned on a mysterious emergent bias inside a model. Each turned on a rule, an input or a step around the score.
Can AI do insurance underwriting?
Yes. AI in insurance underwriting is already on the enforcement record, and the GEICO action shows what it looks like. A tool with AI features picked a new customer for review in a routine 60-day check [source].
The harm came after the flag. She believed she had sent the documents, was not told otherwise, and drove uninsured after her policy was cancelled [source].
The remedy was procedural: a week more to submit documents, one proof of residency rather than two, and adherence to the Pennsylvania Insurance Department’s AI guidance [source].
What the four actions targeted
| Case | What the AI did | What was targeted | Outcome |
|---|---|---|---|
| Earnest | Lending eligibility and pricing | Schools variable, immigration knockout rule, no disparate-impact testing | $2.5M settlement, both rules dropped |
| GEICO | Picked policyholders for underwriting review | Cancellation process and notice | No admission, process changes |
| Apple Card | Credit limits and approvals | Alleged sex bias in the model | No unlawful discrimination found |
| State Farm | Antifraud claims triage | Alleged racial disparity in scrutiny | FHA 3604(b) claim survives dismissal |
Sources for each row are in the case file linked above it.
What to ask an AI underwriting vendor
Start with the record, not the accuracy slide. Every question below maps to something a regulator named in one of the four actions.
Ask for the input list, and flag any proxy tied to where someone studied or lives. Ask whether any rule auto-denies before the model scores. Ask for the last disparate-impact test, with date and method. Ask what the customer sees when the system pulls them for review, and how long they have to answer. Ask how fast an applicant can appeal.
A vendor who can answer the model question but not the notice question has sold you the half of the problem this record shows regulators acting on least.
The bottom line
AI underwriting gets judged on the decision and everything wrapped around it, and the wrapping is where the actions landed. The transferable rule: audit the inputs, the hard rules and the customer’s path through a flag before you audit the weights.
GEICO’s negotiated fix was about a notice and a deadline, not a score.
Sources
- 01Massachusetts Attorney General’s Office, “AG Campbell Announces $2.5 Million Settlement With Student Loan Lender For Unlawful Practices Through AI Use, Other Consumer Protection Violations”, 10 July 2025, https://www.mass.gov/news/ag-campbell-announces-25-million-settlement-with-student-loan-lender-for-unlawful-practices-through-ai-use-other-consumer-protection-violations
- 02Pennsylvania Office of Attorney General, “AG Sunday and GEICO Agree on Improvements to Remedy Customer Confusion and Prevent Unfair Auto Insurance Cancellations”, 22 May 2026, https://www.attorneygeneral.gov/taking-action/ag-sunday-and-geico-agree-on-improvements-to-remedy-customer-confusion-and-prevent-unfair-auto-insurance-cancellations/
- 03Insurance Journal, “Geico to Modify Cancellation Process That Uses AI”, 22 June 2026, https://www.insurancejournal.com/magazines/mag-features/2026/06/22/874424.htm
- 04New York State Department of Financial Services, “DFS Issues Findings on the Apple Card and Its Underwriter Goldman Sachs Bank”, 23 March 2021, https://www.dfs.ny.gov/reports_and_publications/press_releases/pr202103231
- 05New York State Department of Financial Services, “Report on Apple Card Investigation”, March 2021, https://www.dfs.ny.gov/system/files/documents/2021/03/rpt_202103_apple_card_investigation.pdf
- 06U.S. District Court for the Northern District of Illinois, Huskey v. State Farm Fire & Casualty Co., Memorandum Opinion and Order, 11 September 2023, https://storage.courtlistener.com/recap/gov.uscourts.ilnd.424467/gov.uscourts.ilnd.424467.52.0.pdf
Questions
What is AI underwriting?
AI underwriting is the use of algorithmic or machine-learning models to decide whether to accept an applicant and on what terms, for a loan or an insurance policy. In the public record it covers eligibility, pricing and the selection of customers for further review.
Can AI do insurance underwriting?
Yes, AI in insurance underwriting is already on the enforcement record. Pennsylvania's Attorney General said GEICO used a tool with artificial intelligence features to select a new policyholder for further review and underwriting, which led to a cancellation she did not realise had happened.
Has a regulator found AI underwriting discriminatory?
Yes and no: regulators have acted on AI underwriting, but the one model tested at scale was cleared. New York's DFS found no unlawful discrimination in Apple Card underwriting, while Massachusetts settled with Earnest for $2.5 million over a schools variable and an immigration knockout rule, without a court merits finding.
This is analysis, not a verified outcome. It carries no verification badge and never will. The proof lives in the case files, where every figure is checked against the public record and the method is printed on the page.