Why Apple Card’s lawful credit limits still broke trust

In November 2019, complaints on Twitter accused Apple Card’s underwriting of sex bias. New York’s regulator later found no fair lending violation. What failed was explanation: to customers, by the front line, within the law’s narrow notice rules, across a brand-lender partnership.

Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.

Two-minute summaryThe dek and the first sentence of each section, quoted from the article

In November 2019, complaints on Twitter accused Apple Card’s underwriting of sex bias. New York’s regulator later found no fair lending violation. What failed was explanation: to customers, by the front line, within the law’s narrow notice rules, across a brand-lender partnership.

  1. The week Apple and Goldman Sachs launched Apple Card in August 2019.
  2. The decision record: what the bank could reconstruct The part of the system that drew the most suspicion held up best under examination.
  3. The front line: where the explanation had to happen The complaint that ignited the story was about a conversation as much as a credit limit.
  4. The legal notice: what the rules required, and where they stopped Compliance teams define their explanation duties by the legal trigger, and the trigger here is adverse action.
  5. The partnership: whose name was on the decision The fourth place where explanation had to work was between two companies.
  6. The strongest objections, and where they hold Three objections deserve a fair hearing.
  7. With hindsight DFS published its findings on March 23, 2021.

The week

Apple and Goldman Sachs launched Apple Card in August 2019. The card carried Apple’s name, but the credit came from Goldman Sachs Bank USA, a New York state-chartered bank. According to the later report of the New York State Department of Financial Services (DFS), the bank was responsible for the credit policy, the underwriting decisions, the management of the program and customer service. According to the Consumer Financial Protection Bureau (CFPB), Apple designed the screens customers used to manage their accounts and the creative approach of the advertising.

The pitch was speed and simplicity. Applicants could apply on an iPhone and, as Apple put it in its marketing, start using the card in minutes. Advertising said the card was created by Apple, not a bank. DFS later recorded that millions of Americans applied in the first few months, including hundreds of thousands of New Yorkers. One design choice stood out: the bank ran a hard inquiry on an applicant’s credit file only after an offer had been made and accepted, so a declined application did not lower the applicant’s score.

On November 7, 2019, a technology entrepreneur began a series of posts on Twitter about his household’s offers. He and his wife filed joint tax returns and lived in a community property state, yet he said his Apple Card credit limit was 20 times hers. He attributed the gap to sex discrimination. He suggested that the bank relied on algorithms and machine learning, and he complained that a customer service agent could not explain those algorithms or the reason for the difference. He also objected to a policy that made new account holders wait six months before they could seek a higher limit.

The posts drew a large response, including posts from a co-founder of Apple. He said that he too had been offered far better terms than his wife, although the couple held their accounts and assets jointly. Within days, according to DFS, the bank re-reviewed both women’s credit files and raised their limits to match their husbands’. At the same time it dropped the six-month waiting period for appeals of credit decisions.

DFS is New York’s financial services and insurance regulator. By its own description it supervises about 1,500 banking and other financial institutions, and it enforces federal and New York fair lending law. It launched an investigation into whether the bank had discriminated against women in Apple Card underwriting. Consumers who believed they had been treated unfairly began sending complaints to the Department, many of them surprised by a denial or by terms less generous than they expected.

The legal frame was older than the product. The Equal Credit Opportunity Act and its implementing rule, Regulation B, bar creditors from considering an applicant’s sex or marital status. When a creditor takes adverse action, such as a denial, the applicant is entitled to a statement of specific reasons. Regulation B also generally bars a creditor from asking an applicant’s sex at all, outside a formal self-test. An applicant who is approved, on terms they never specifically requested, generally receives no statement of reasons. That gap between what the law requires a lender to explain and what a surprised customer wants explained is where the week’s story unfolded.

The Apple Card episode began as a public accusation that an algorithm was biased. The regulator’s eventual finding pointed the other way. In March 2021, DFS reported no fair lending violation, and it found that the bank could explain each complainant’s decision when the Department asked. What failed in November 2019 was the bank’s ability to explain its decisions to its own customers, at the moment they were surprised, through the channel they actually used.

That distinction matters to any bank, credit union or card issuer that automates credit decisions. Explanation is a capability that has to work in several separate places, each with a different owner. This article takes them in turn. The first is the decision record: whether the institution can reconstruct why a decision came out as it did. The second is the front line: whether the people who answer the phone can say anything useful and route an appeal. The third is the legal notice: what the rules require a creditor to explain, and where their boundary leaves a customer with nothing. The fourth is the partnership: who speaks for a decision when the brand on the card is not the lender. A fifth section weighs the strongest objections to this reading.

The four places do not overlap. The record is a data and model problem owned by risk and technology. The front line is an operations problem owned by servicing. The notice is a legal problem owned by compliance. The partnership is a contract and governance problem owned by the executives who sign the deal. An institution can be strong in one and fail in another, which is roughly what happened here.

The decision record: what the bank could reconstruct

The part of the system that drew the most suspicion held up best under examination. According to the DFS report, the Department asked the bank to explain the credit decision for each person who complained to it. In each case the bank identified the factors behind the decision. They were ordinary credit factors: credit score, indebtedness, income, credit utilization, missed payments and other elements of credit history. DFS found the decisions consistent with the bank’s credit policy, and none of the factors was an unlawful basis for a decision.

DFS also tested the portfolio as a whole. Its Consumer Examinations Unit ran a regression analysis on underwriting data for nearly 400,000 New York applicants, covering applications from the card’s launch until the first complaints. The analysis did not show a disparate impact on women. The Department also found that the bank had a fair lending program designed to make sure its credit policy and the underlying statistical model did not consider prohibited characteristics.

Why a reconstruction is harder than it sounds

Being able to say why one applicant got one limit, months after the fact, takes more than keeping the model. An institution needs at least four things preserved for every decision. It needs the inputs as they stood at that moment, including the credit bureau data pulled on that day, which can change week to week. It needs the version of the model that scored the application. It needs the version of the policy rules that sat on top of the model, such as minimum limits, caps and exclusions. And it needs the ranking of factors that drove the result, computed in a way that can be checked.

Each of those has an owner and a failure mode. Bureau data may be retained only in summary form. Model versions may be overwritten when a new one is deployed. Policy overlays may live in a separate decision engine configured by a different team. Factor rankings may be produced by a separate explanation tool that approximates the model rather than reading it. If any one of these is missing, the institution can describe its model in general terms but cannot explain a particular customer’s outcome.

How long the record has to last

Regulation B sets a floor for retention. For consumer credit, a creditor must keep the application and the information used to evaluate it. It must also keep any notice of action taken and any statement of reasons. The period is 25 months after notifying the applicant of the action. It must also keep any written statement in which the applicant alleges a violation. Those are minimums for documents.

For an automated decision, the documents are only part of what an explanation needs. The model version, the policy rules in force and the factor outputs are what let anyone say why. Retaining them for at least as long as the applications themselves is a sensible design rule, and one that costs little once it is built into the decision engine. Rebuilding them after the fact may be impossible.

Explanation tools and their limits

An institution that uses a complex model may produce reasons with a separate explanation method that estimates how much each input contributed to a score. Such methods can be useful, but, as the CFPB noted in 2022, they approximate the model rather than reading it directly. Two applicants with similar files can receive different top reasons if the method is unstable. A reason can also appear that the policy layer, not the model, actually drove.

Validation therefore has to cover the explanation as well as the prediction. A sound program checks that reasons are stable across small changes in inputs. It checks that they agree with the factors the model and policy actually used. And it checks that they map to statements a customer can understand. It also checks that the same reason is produced whether the decision is pulled for a notice, for a service agent or for an examiner. If three systems compute reasons three ways, the institution has three explanations and no single record.

Testing for sex without a sex field

The fair lending test raises its own data problem. Regulation B generally prohibits a creditor from asking about an applicant’s sex in connection with a credit transaction. The main exception is a self-test that meets the rule’s conditions, including telling the applicant that providing the information is voluntary. A card issuer that wants to know whether its outcomes differ by sex therefore may lack a sex field to test against. It must either run a formal self-test or estimate sex by other means, and each route has limits on accuracy and on how results can be used.

The DFS report does not describe how the Department’s analysts handled this. The general point stands for any lender. Fair lending assurance for an automated product is a standing analytical program, built before launch, with a method the institution can defend. It cannot be assembled in the week a complaint goes viral.

Lawful inputs, inherited patterns

The DFS report made a broader point that applies to any automated underwriting. Even when credit scoring complies with the law, it can reflect and carry forward older inequality. Income, assets and credit history are lawful factors, but they are shaped by the country’s history of discrimination. The report also noted that lenders test new models on past lending data, which may carry bias from earlier products.

For a lender, this means outcome testing cannot stop once the inputs are confirmed lawful. A model can use only permitted variables and still produce patterns that a disparate impact analysis would flag. A sound fair lending program tests outcomes on a schedule and investigates differences it finds. A program can also record whether a less discriminatory alternative was considered, which is a practice rather than a requirement cited here. Organizations should check with counsel on the legal standard that applies in their jurisdictions, particularly since the 2026 changes to Regulation B described in the hindsight section.

Why spouses got different answers

The report also explains the specific puzzle that started the story. DFS found that couples who said they shared finances typically had different credit scores, and that a closer look showed differences in their credit profiles. In some cases one spouse was named on the mortgage and the other was not. In others one spouse carried several cards and a line of credit while the other held a single card. A longer, more varied credit history can produce better terms than a thin one, even when both are paid on time. The Apple Card application asked for all sources of income, including income from someone else, but the credit files themselves were individual.

A sophisticated model was not needed to produce these differences. A traditional scorecard would have treated the two files differently too. For that reason the decision record, though necessary, did not settle the public question. The bank could explain the outcome to a regulator. It had not prepared to explain it to a spouse.

What an examiner would ask to see

For an institution running automated credit decisions today, the decision record is the part an examiner can test directly. A reasonable evidence pack would include model documentation and validation reports, and fair lending testing results with the method described. It would add a data lineage map from bureau pull to decision, and change logs for the model and policy rules. We would put first a demonstrated ability to pull any single applicant’s decision and its reasons on request, and in a time the examiner considers reasonable.

The differences by type of institution are real. A large issuer may build all of this in house. A credit union or community bank that uses a vendor’s scoring model may find that the decision record lives partly at the vendor. In that case the contract has to give the institution access to the inputs, versions and factor outputs for its own decisions. Without that clause, the institution remains responsible for decisions it cannot reconstruct.

The front line: where the explanation had to happen

The complaint that ignited the story was about a conversation as much as a credit limit. The customer said that the service agent could not explain the algorithm or the basis for the difference between his limit and his wife’s. That is a front-line failure, and it is different in kind from a model failure.

The DFS report was direct about it. The Department found that lack of transparency to the complainants seemed to produce confusion that could have been reduced. It said the bank seemed unprepared for the possibility that a complex underwriting model that let applicants start using the card right away would produce outcomes that surprised them. And it noted that internal deadlines and pressure to roll out the card by a particular date compounded these issues.

The six-month wait

The six-month waiting period shows how a servicing rule can turn a surprise into a grievance. A customer who thought a limit was wrong had no route to have it looked at again for half a year. DFS observed that the bank’s quick change of policy, in the aftermath of the media scrutiny, showed it was capable of reviewing credit terms faster than it had first allowed. The delay was a policy setting, and policy settings can be changed in days.

It is natural to design appeal paths around denials, because denials are where the legal duties sit. A product that approves almost everyone at some limit moves the surprise from denial to terms. The appeal path has to move with it. For an approved customer, a credible appeal means a named route, a re-review by someone with authority to change the outcome, and an answer in days.

What an agent needs on screen

A service agent cannot explain a model. An agent can explain the main factors in a decision, if the system gives them those factors in language the agent can repeat. That requires the decision record from the previous section to flow into the servicing system, translated into short reason statements. It also requires agreement on how much to say. The reasons an agent gives must match any written notice the customer receives, or the institution creates a second, inconsistent explanation.

There are trade-offs. Putting credit report details in front of agents widens access to sensitive data and raises training and monitoring needs. Scripted explanations can sound evasive if they are too generic. Unscripted ones can drift from the factors actually used. A middle path is a small set of approved reason statements, generated from the decision record, with an escalation path to a credit specialist for anything beyond them.

Routing complaints that allege discrimination

Not every complaint about a credit limit is a fair lending complaint, but some are, and the front line has to tell them apart. A complaint that says a decision was unfair because of sex, marital status or another protected characteristic should go to fair lending review as well as to servicing. That routing needs a tag in the complaint system, a rule for applying it and someone who reads the tagged cases as a set.

The set matters more than any single case. One complaint about a spouse’s limit says little. Dozens in a week, arriving through a partner’s channel and on social media at the same time, are a signal that the fair lending team should be looking at the portfolio data immediately. An institution with good routing would see such a pattern in its own data as it formed, whatever was happening in public.

Product rules that create complaints

Some of the front-line pressure came from product design rather than underwriting. Apple Card did not allow authorized users. DFS noted that authorized user status is frequently used by couples who share finances. Without it, each spouse who wanted the card had to apply and be underwritten separately, which made differences between spouses visible in a way a shared account would have hidden. Consumer frustration with that policy, DFS said, drew attention to a misconception: someone who relies on a spouse’s credit as an authorized user may wrongly believe they have the same credit profile.

Nobody in a model risk function would normally review a decision about authorized users. Yet that decision shaped how many surprised customers the front line would face. A launch review for an automated credit product should include servicing and product policy alongside the model, because the three together determine how many explanations will be needed.

Measures an operations leader can track

Front-line readiness can be measured. Useful measures include the share of complaints that mention an unexplained decision, the time from appeal to answer and the share of appeals that change the outcome. Quality reviews can also score whether agent explanations match notice reasons. A high reversal rate on appeal is a signal about the model or policy. A low rate combined with many complaints is a signal about explanation.

Compliance teams define their explanation duties by the legal trigger, and the trigger here is adverse action. Regulation B defines it in three parts. It covers a refusal to grant credit in substantially the amount or on substantially the terms requested, unless the creditor makes a counteroffer that the applicant uses or accepts. It covers a termination or an unfavorable change in the terms of an account that does not affect all or substantially all of a class of accounts. And it covers a refusal to increase the credit available to an applicant who has applied for an increase.

When adverse action is taken, the creditor must give a written notice. The notice must contain either a statement of specific reasons or a disclosure of the applicant’s right to ask for one. The reasons must be specific and give the principal reasons for the action. Regulation B says that statements that the applicant failed to reach a qualifying score, or did not meet internal standards, are not enough.

The gap for approved applicants

An applicant for a card may not request a specific credit limit at all. If the issuer approves the application at a limit lower than the applicant hoped, there may be no refusal of an amount requested, and so no adverse action and no statement of reasons. The DFS report made the same observation: federal law requires lenders to explain credit denials, but not the reasons for the amount and terms of credit granted.

The six-month waiting period interacted with this rule in a way that is easy to miss. A refusal to increase a limit, when the customer has applied for an increase, is adverse action. By postponing the point at which a customer could make that request, the waiting period also postponed the point at which the customer would have been owed reasons. Whether or not anyone intended that effect, it lengthened the period in which a surprised customer had neither an explanation nor a right to one.

Counteroffers and requested amounts

The definition has a nuance that matters for product design. If an application asks the applicant to name a desired limit, and the creditor grants less, the creditor has refused credit in the amount requested. Unless the applicant uses or accepts the lower amount as a counteroffer, that refusal is adverse action. Regulation B also sets timing. Notice of action on a completed application is generally due within 30 days. Where the applicant does not accept or use a counteroffer, notice is due within 90 days after the counteroffer.

So the design of the application form shapes the legal duties. A form that never asks for an amount produces fewer adverse actions on terms. A form that does ask creates more notices but also more explanations. Neither is wrong. The choice should be made deliberately by product, compliance and servicing together, with the explanation consequences in view.

State law adds its own layer

Federal law was only part of the frame. The DFS report notes that the New York State Human Rights Law has its own equal credit provisions. They prohibit discrimination on the basis of race, creed, color, national origin, sexual orientation, gender identity or expression, military status, age, sex, marital status, disability or familial status. That list is longer than the federal one. A lender operating in several states has to test against the broadest set of protected characteristics that applies to each portfolio, which affects both the testing method and the data it needs.

Reasons must match the model

The law also constrains what reasons may say. The official interpretations published with Regulation B require that the reasons relate to and accurately describe the factors actually considered or scored. A creditor may not tick the closest item on a sample form if that item was not the real factor. The Fair Credit Reporting Act separately requires disclosure of key factors that affected a credit score. The same interpretations say that disclosure does not by itself satisfy the duty to give reasons for the creditor’s own decision.

For a model with many inputs and interactions, producing accurate principal reasons is a design requirement, not an afterthought. It has to be built into model selection and validation. A model whose reasons cannot be produced reliably creates legal risk at the denial stage even if it never draws a viral complaint at the approval stage.

Alternative data and the second look

The DFS report describes another feature of the program that bears on notices and data. In the months after launch, the bank ran a second-look program for some applicants with limited or no credit history who would otherwise have been declined. Those applicants could choose to have Apple release their history of buying Apple devices and the payment methods on file to the bank. Applicants approved this way received a limit of at most $2,500. The bank ended the program after concluding that the data did not improve underwriting accuracy. DFS found no violation associated with it.

The report also notes that in December 2019 several federal agencies issued an interagency statement on the use of alternative data in underwriting. That statement said such data might improve the speed and accuracy of decisions. It also said fair lending law applies to its use and urged appropriate testing, monitoring and controls. For compliance teams, the practical consequence is that each new data source becomes a new potential reason, which has to be explainable in a notice if it drives a denial.

Deciding to explain more than the law requires

An institution can choose to explain approvals as well as denials. The DFS report noted that the Apple Card website later added a page describing the data the bank draws on in setting terms. It listed utility payment histories, disposable income and TransUnion credit report information, and explained how applicants might improve their chances. That is a voluntary explanation at the level of factors, not a disclosure of the model. It is a reasonable template for any lender whose automated decisions set terms as well as approve or deny. Organizations considering such a policy should ask counsel whether voluntary explanations could affect their notice obligations.

The partnership: whose name was on the decision

The fourth place where explanation had to work was between two companies. Apple’s brand, devices and interface were what customers saw. Goldman Sachs Bank USA made the decisions and owned the legal duties. When a customer wanted an explanation, they asked the brand they knew, through the interface Apple designed, and the answer had to come from the bank.

Expectations set by one party, duties held by another

The DFS report records the marketing claims. The card was advertised as created by Apple, not a bank, and as helping users lead a healthier financial life. Apple’s chief executive said, in announcing it, that it was designed to put privacy and security first. Those claims set customer expectations about simplicity and fairness. None of them was the bank’s to make, yet the bank carried the regulatory consequences when customers felt those expectations were not met.

The same split arises whenever a brand and a chartered lender divide a product. A credit union may offer a card through a fintech’s app. A community bank may lend through a partner’s checkout flow. In each case the institution with the charter owns the credit decision and its notices, while the partner owns the screen, the brand and possibly the first customer conversation.

What the contract has to settle before launch

A partnership agreement for a credit product should answer several questions explicitly. Who drafts the customer-facing explanation of credit decisions, and who approves it. Who sees complaint data, in what detail and how fast. Who can change credit policy or servicing rules, and how quickly, when a problem appears. Where in the partner’s interface an explanation or appeal route appears, and who controls that screen. And who speaks publicly when a decision is questioned.

The episode shows why speed of change matters. The bank changed the appeal policy within days. If a partner had controlled the relevant screens or scripts and needed a long release cycle, the fix would have taken longer. Launch readiness also needs a joint gate. The CFPB later found that the bank’s board was warned before launch that the dispute system was not fully ready, a matter covered in the hindsight below. A partnership that runs to a launch date set by the brand partner puts pressure on exactly the controls that protect the chartered institution.

Complaints as shared data

In a brand-lender partnership, complaints arrive through several doors. Some go to the lender’s service line. Some go to the brand partner’s support channels. Some go to regulators, as many did to DFS in this case. Some appear only on social media. No single party sees them all unless the contract and the systems require it.

A workable arrangement sends every complaint about a credit decision, wherever it arrives, into a shared queue the chartered institution can see. Each complaint should carry the same tags, so the fair lending and servicing teams can count patterns across channels. The partner should also commit to escalation times for anything that alleges discrimination. Without those terms, the institution that holds the legal duty may be the last to learn about the problem.

A launch gate both parties can stop

The episode argues for a joint launch gate with clear stopping power. Before launch, both parties would confirm that servicing scripts, appeal routes, complaint routing and dispute handling work at the expected volume as well as in testing. The chartered institution’s risk and compliance functions should be able to delay launch without the decision being treated as a breach of the partnership.

That is a governance choice, and it is uncomfortable when a launch date has been announced. The alternative is that the institution with the legal duties launches with known gaps. Boards of institutions entering such partnerships can ask a simple question before any launch: which readiness items are incomplete, and who decided to proceed anyway.

Remediation as a joint product

The response to the episode was itself a joint effort. DFS reported that, during the investigation, the bank and Apple launched Path to Apple Card in June 2020. It gave declined applicants step-by-step instructions to improve their credit, with approval as the goal. DFS reported that more than 70,000 consumers had enrolled and nearly 5,000 had been approved by the time of its report. Whatever one thinks of the program’s design, it shows that the remedy for an explanation failure sat across both companies.

The strongest objections, and where they hold

Three objections deserve a fair hearing.

No violation, so a public relations problem

The first is that DFS found no fair lending violation, so this was a public relations problem rather than a risk management one. The finding is real, and it matters. But the same report criticized customer service and transparency, and the bank and Apple changed policies and, during the investigation, launched a remediation program. A regulator’s investigation, policy changes and a joint program carry real cost. An institution that treats explanation as public relations will budget for it as public relations, which means after the fact.

Explanations invite gaming

The second objection is that explaining approved terms invites applicants to game the model or exposes proprietary methods. Explanations at the level of factors, of the kind the Apple Card website later published, reveal little that is not already in consumer guidance. Regulation B already requires factor-level reasons for adverse action. Extending the same reasons to surprised approved customers adds little new exposure.

No script fixes a misconception

The third objection is stronger. Many complaints rested on a misunderstanding about how household finances relate to individual credit files, and no script would have dispelled every one of them. DFS devoted a section of its report to exactly these misconceptions. The objection is partly right: some customers would have stayed unhappy with any answer. But the agent in the first complaint could not explain the basis for the difference, so the misconception had nothing to compete with. A short, accurate account of the factors in each spouse’s file would not have satisfied everyone. It would have given the bank something true to say on the first day rather than in a report published in March 2021.

Regulators should set the standard

A fourth objection is that lenders should not be expected to build explanation capabilities beyond what regulators require. If customers deserve explanations of approved terms, the argument runs, the rule should say so. There is something to this. A clear standard would let every lender build to the same target and would remove the competitive cost of explaining more than rivals do.

But the absence of a rule did not protect the bank here. The reputational and supervisory consequences arrived without one. Federal guidance on explanation has also shifted since 2019, as the hindsight section shows. An institution that waits for a settled federal standard may wait a long time. Building to the customer’s reasonable expectation, within what counsel confirms is consistent with the rules, is the more stable plan.

Where the argument is weakest is with models that are genuinely hard to interpret. If an institution cannot produce accurate principal reasons from its model, better scripts will not help. The answer there sits back in the decision record: choosing and validating models with explanation as a requirement from the start.

With hindsight

DFS published its findings on March 23, 2021. It reported that its review of several thousand pages of records, interviews and underwriting data for about 400,000 New York applicants did not produce evidence of unlawful discrimination. It found that women and men with similar credit characteristics generally had similar outcomes, and that decisions for complainants were explainable, lawful and consistent with the bank’s policy. It also concluded that deficiencies in customer service and a perceived lack of transparency had undermined consumer trust. The report went further than the case, arguing that credit scoring and anti-discrimination law needed strengthening and modernizing to improve access to credit.

Federal guidance then moved toward the problem this article describes, and later away from it. On May 26, 2022, the CFPB issued Circular 2022-03. It said that the adverse action requirements of the Equal Credit Opportunity Act and Regulation B apply to all credit decisions regardless of the technology used. Creditors may not use complex algorithms when doing so means they cannot give specific and accurate reasons. A creditor’s lack of understanding of its own methods is no defense. The circular added that post-hoc explanation methods only approximate models, and creditors must be able to validate those approximations. A later circular, numbered 2023-03, addressed adverse action notices and the proper use of the sample forms in Regulation B.

On October 23, 2024, the CFPB issued orders against Goldman Sachs Bank USA and Apple over Apple Card. The orders did not concern underwriting. The Bureau found that the bank’s board had been warned on August 16, 2019 that the card’s dispute system was not fully ready, and that the card launched on August 20. It found violations of the Truth in Lending Act and Regulation Z in how disputes were handled, an unfair practice in delays, and deceptive practices around Apple Card Monthly Installments. Goldman was ordered to pay $19.8 million in redress and a $45 million civil money penalty, and Apple a $25 million penalty. As checked on November 14, 2026, the CFPB’s page listed the matter as post order.

On May 12, 2025, the CFPB withdrew a long list of guidance documents, including both Circular 2022-03 and Circular 2023-03. Its notice said that the Bureau intended to avoid issuing guidance except where necessary and that it was reducing its enforcement activities. The withdrawal did not change Regulation B itself. As in effect on November 1, 2026, the rule still defines adverse action as before and still requires a statement of specific reasons or notice of the right to one.

Regulation B did change in another respect. Amendments published in 2026 added text stating that the Act “does not provide that the ‘effects test’ applies for determining whether there is discrimination in violation of the Act.” The official interpretations now say that the Act “does not provide for the prohibition of practices that are facially neutral as to prohibited bases, except to the extent that facially neutral criteria function as proxies for protected characteristics designed or applied with the intention of advantaging or disadvantaging individuals based on protected characteristics.” How much federal footing remains for disparate impact testing of the kind DFS described is a question for counsel.

What held up is the operational argument. The 2021 report and the 2024 orders tell the same story from two angles. A product launched to a date, with servicing and explanation capabilities that were not ready, under a partnership that split brand from responsibility. What did not hold up is any expectation that federal guidance would settle how explanation should work for complex models. The guidance was issued in 2022 and withdrawn in 2025. The underlying duty to give accurate, specific reasons stayed in the regulation, and the operational work of meeting it remained with each lender.

Spotted an error? Tell us and we will check it and correct the article where needed.

  • Apple Card
  • fair lending
  • Regulation B
  • adverse action
  • explainability
  • NYDFS