Upstart rejected its fair-lending monitor's less-discriminatory model, ending the monitorship

An independent fair lending monitor (Relman Colfax) found statistically significant approval disparities for Black applicants in Upstart's AI lending model during a multi-year oversight process from December 2020 through March 2024. The monitor proposed a less discriminatory alternative (LDA) model to address these disparities, but Upstart rejected it on accuracy grounds and offered its own alternative, which the monitor declined to validate. The disagreement ended the monitorship in an impasse, leaving the approval disparities unremediated.

Upstart · Incident Mar 27, 2024 · Indexed Jun 4, 2026 · 3 sources

Upstart chose model accuracy over the monitor's less discriminatory alternative, leaving confirmed approval disparities for Black borrowers unremedied.
What
An independent fair lending monitor (Relman Colfax) found statistically significant approval disparities for Black applicants in Upstart's AI lending model during a multi-year oversight process from December 2020 through March 2024.
Incident date
Mar 27, 2024
Who
Upstart
Failure mode
Policy Violation
AI surface
Agentic Workflow
Severity
High

What happened

From December 2020 through March 2024, the civil rights law firm Relman Colfax served as an independent fair lending monitor over Upstart's AI lending model. The monitor identified statistically significant approval disparities for Black applicants, though it found no pricing disparities or close proxy variables for protected classes. The monitor proposed a less discriminatory alternative (LDA) model to address the approval disparities, but Upstart rejected it, claiming it would unacceptably compromise model accuracy. Upstart instead proposed its own alternative LDA, which the monitor did not validate, leading to an impasse that ended the monitorship without resolving the identified approval disparities. The parties jointly warned of the risk of AI-driven bias in consumer lending.

What broke inside the model

Failure path · mode profile · Policy Violation
  1. 01 · TriggerA prompt pushes against a deployment boundary.
  2. 02 · Model stepThe model produces the disallowed output.
  3. 03 · Control gapNo enforcement blocks it at generation time.
  4. 04 · FailureThe output crosses the policy line.
  5. 05 · ConsequenceA limit the business set is breached in public.

The output crosses a policy boundary the deployment had defined.

Upstart's AI lending model used educational data inputs (degree attainment, program of study, school attended) that were closely correlated with race, producing approval rate disparities for Black applicants even though no individual variable operated as a close proxy for protected classes. The company's prioritization of model predictive accuracy over the monitor's proposed less discriminatory alternative created a fundamental tension between algorithmic performance and fair lending compliance that the oversight process could not resolve.

Public visibilityMedium
Regulatory exposurePossible
Customer impactClass-wide
Financial impactUnknown
Time to disclosureMonths
  1. PrimaryLDF, SBPC, and Upstart Announce Final Monitorship Report on AI and Fair Lendingnaacpldf.org
  2. Customer-DisclosedLDF, SBPC, and Upstart Announce Final Monitorship Report on AI and Fair Lending | Upstart Holdings, Inc.ir.upstart.com
  3. PrimaryFair Lending Monitorship of Upstart Network's Lending Model: Relman Colfax PLLCrelmanlaw.com
Permalinkhttps://failureindex.ai/failures/upstart-rejected-fair-lending-monitor-less
CitationAI Failure Index. "Upstart rejected its fair-lending monitor's less-discriminatory model, ending the monitorship" (FI-0088). Realm Labs. https://failureindex.ai/failures/upstart-rejected-fair-lending-monitor-less (indexed Jun 4, 2026).
Share cardA branded image of this record for posts and slides.

Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0088. Full dataset at /data.

Note from Realm Labs, the Index steward

How Realm would have caught this

Controls for this failure mode
  • Prism
  • OmniGuard

Realm compares what the model is about to output or do against the policy that governs the deployment, in real time, and can deny or redact the action before it takes effect, which is the gap an after-the-fact review never closes in time.