Dave ExtraCash (CashAI)

What this models

This example runs on the Cash-advance-app-class amount setting behind an advertised maximum network: 11 components and 24 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 7 assumed · 6 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• assumed

  D48-derived new org (Phase 6, lending-credit-collections), re-derived at the coarsest granularity its record supports. TOPOLOGY. Eleven nodes, all documented, none decorative, and the shape's defining feature is that the deployment's automated surfaces and its two review channels never meet. TWO MODELS carry the THREE automated surfaces the operator discloses in its own annual report: the engine that sets eligibility and amount, drawn together with the real-time behavioural fraud system and its user-level controls because that system reads the same account, gates the same determination, is recorded against the same origination and answers to the same function; and the income-and-expense prediction component, drawn on its own because the operator states it feeds that decision as well as the member-facing budgeting view. ONE INPUT SOURCE carries the two external collections described — the linked checking-account transaction history both models read, and the device and session behaviour the fraud system reads. THREE RECORD STORES: the origination and repayment history the engine retrains from, the support contact and cancellation record, and the advertised terms and fee disclosure, which is the only store in the deployment an outsider can read. THREE OPERATOR CLASSES: the risk and model function that owns every automated surface, the product and growth function that owns the presentation and the experiments, and the support staff who own neither. TWO REVIEWERS, both external and both wired in on documented reads: the partner banks that originate the advances, and the consumer-protection enforcement channel. NO enforcement node, and the reason is evidential rather than stylistic — the defendants expressly deny that they collected on advances, so drawing a downstream collections system would put a contested characterization on the diagram. NO worklist and NO retriever: no queue, backlog, triage list or retrieval step appears anywhere in this record. NO guardrail: the fraud system gates whether an advance proceeds and is drawn as part of the automated determination the operator says it is, not as a bounded check over the amount decision. NO externalBoundary and no egress pathway: no data broker, ungoverned host, unsanctioned tool, third-party disclosure or breach appears anywhere in this record, and no privacy count is pleaded.

• baseline

  DRAWN AT THE COARSEST GRANULARITY THE RECORD SUPPORTS. This board was first derived at thirteen parts and forty-two pathways and has been re-derived at eleven and twenty-four without dropping a documented fact: every flow the record describes is still on it, drawn once, and every part the case, its containment or an instrument the record documents turns on is still drawn separately. The fraud system is narrated on the engine it runs beside, and the session signals it reads on the account input, because every pathway the fraud system had ran beside an engine pathway between the same parties and nothing in the case touches it. Where one documented flow had been drawn twice, the survivor now carries the rest in its own words. The amount reaching support is narrated on the low offer becoming a contact, which is how the record says it arrives. The engine's aggregate behaviour and the fraud output reaching the model function are narrated on that function's read of the matured file. The fee structure reaching the engine is narrated on the revenue objective travelling to the function that tunes it. Complaint content failing to reach that function is narrated on support's own channel to it. The product class reading its own terms back, model versions stamped on the file, the bank booking the origination and seeing the file, and the projection being retained, rendered in the budgeting view and maintained by the model function are each narrated on the part or pathway they belong to. The contacts-against-terms reconciliation is narrated on the reconciliation of the terms against the file and on the proceeding reaching product, because the record ties the one documented repair to the litigation. A correction to the terms ordered from outside is narrated on the enforcement channel: no order has issued. No width moved in the process.

• baseline

  THE MEMBER IS NOT ON THIS DIAGRAM, AND THE ABSENCE IS THE DEPLOYMENT'S SHAPE RATHER THAN A MODELLING SIMPLIFICATION. Served people are never modelled as dynamics anywhere in this catalogue, and here that boundary lands on the one party who sees the output: the member is the data subject, the decision subject, and the only party ever shown the amount. They are shown no reason for it, there is no adverse-action explanation, there is no appeal from an amount, and the operator's own description of the process is fully automated. So the widest consumption of this engine's output leaves the network entirely, and every pathway drawn from the engine to a person is a pathway to someone who is not the person the number is about. That is why the board's two operator-facing engine pathways read 2 and 1 rather than 3: the strong one is not on the diagram, because it is not an operator. The engine's output reaches the model function through the file it writes, and reaches support as a contact about the amount, and both are drawn through those records rather than straight from the engine.

• baseline

  WHERE THE LAB SHAPE DIVERGES FROM THE PAN SHAPE, in four places, each stated rather than smoothed. (1) PAN carries the linked checking-account feed as a STORE; the Lab draws it as an inputSource, because it is the member's own external account read by the models rather than a record this deployment authors, and §5.1 licenses the distinction. The behavioural and device signals ride that same input on both sides. (2) PAN carries the fraud system as a model of its own; the Lab draws it inside the amount engine, because every pathway it has runs beside an engine pathway between the same parties and nothing in the case turns on it separately. (3) PAN carries the Federal Trade Commission, the Department of Justice and the court as governance ACTORS rather than as a user class; the Lab draws one enforcement reviewer, because a channel that read the published page and the internal record and pressed the class that owns the terms needs a node to act from. (4) PAN has no edge kind for a check and no model-to-model edge kind at all, so three PAN peer edges are redrawn here as Lab checks and three model couplings are Lab-side entirely; the redrawn three keep their PAN widths on the stated rung mapping. Nothing is asserted on the Lab side that the PAN file does not already record.

• assumed

  DEMAND 3 / CAPACITY 1. Demand 3 rests entirely on the operator's own SEC-filed reporting: over 19 million members signed up since inception and over 14 million having used at least one product; origination volume rising from approximately $5.1 billion in fiscal 2024 to approximately $7.6 billion in fiscal 2025; average monthly transacting members up approximately 17 per cent year over year; an average advance of $205 on an average term of about eleven days; and a determination remade each time a member opens the home screen or the advance section, against approximately 280 full-time employees. Capacity 1 is a counterfactual rather than a measurement and is labelled as one: the product is a $205 advance authorized in under five minutes with no credit check and no credit-bureau data, and the record documents no per-application human underwriter anywhere in this deployment. It sits above 0 because small-dollar consumer credit is underwritten by people elsewhere in this domain, so a manual floor exists; it sits well below the 2 default because nothing in this record measures this deployment against its own human counterfactual, and no comparator of any kind — peer, historical or internal — appears in it.

• baseline

  BASELINES. The eight pathways at 3 are the ones the operator's own description makes wide: the account feed into the engine and into the projection (its whole decision input, on every determination), the matured-outcome retraining read, the determination write that fills it, the model function reading that file and versioning the engine against it, and the support class reading and writing its own contact record. Their width is the design the operator publishes, not a governance lapse, and reading them as a lapse would misstate the case. The pathways at 2 are the offer distribution and the contact analyses reaching the product class, that class's authorship of the terms, the published page being read from outside, the revenue objective travelling to the model function, and the two model couplings — each anchored to a specific operator statement or pleaded allegation. The pathways at 1 are, without exception, ones the record shows running thin: the amount reaching the bank that extends it, the contact record read that took a civil investigative demand to run, the check arms of both reviewers, support's findings raised to product, a low offer becoming a contact, and the reconciliation of the advertised number against the file. TWO PATHWAYS CARRY NOTHING BEFORE A LEVER OR A PRESSURE MOVES THEM, and each is a documented cut rather than an assumption: an independent second read of the engine, which no model documentation, validation report, fairness assessment or external evaluation in the public record supplies; and the channel from support, and from the complaint content support handles, to the function that sets amounts, which PAN itself scores at the bottom of its own file. A correction to the advertised terms ordered from outside is the third documented absence, and it is narrated on the enforcement channel rather than drawn, because there is no settlement, no consent order and no adjudication on the merits, and every documented change to those terms arrived through the class that authors them.

• assumed

  THE LOAD-BEARING QUANTITY IN THIS CASE IS CONTESTED MEASUREMENT, TWICE OVER, AND NO BASELINE HERE RESTS ON THE DISPUTED FIGURES. The pleading alleges that in the first fourteen months after the operator began advertising up to $500, the maximum was offered to a new user 0.002 per cent of the time — fewer than one determination in forty-five thousand — that fewer than one new user in 750 was offered even half of it, that the most common offer was $25, that more than three-quarters of the time no advance was offered at all, and that on average more than 40 per cent of new users obtained no offer in a calendar month while the monthly subscription was charged regardless. The defendants deny paragraphs 34, 35 and 36 in their entirety, and the court's footnote 2 records that they contend the government's method of calculating the figures is wrong and its data incomplete, a dispute a Rule 12(b)(6) motion cannot resolve. The window itself is never dated in the pleading. So those figures appear in this board's copy as the government's computation from the operator's own data, disputed by the operator, over an unspecified window — and the reconciliation pathway they belong to is drawn from the UNDISPUTED half instead: an advertised maximum of $500 against the operator's own published average of $170 for fiscal 2024 and $205 for fiscal 2025.

• assumed

  THE MODEL LAYER IS OPERATOR-TIER THROUGHOUT AND IS LABELLED SO WHEREVER IT APPEARS. Approval rates at all-time highs, higher average approval amounts, lower delinquency rates, improved risk ranking, the training-set sizes, the feature-count claim and the eleven-day refinement cadence are all Dave's own statements in SEC filings and press releases. No third party has validated the engine, its accuracy, its calibration or its distributional effects; no model documentation, validation report or fairness assessment exists in the public record; and the words algorithm, artificial intelligence, machine learning, model and underwriting appear nowhere in the original complaint, the amended complaint, the dismissal memorandum, the government's opposition, the answer, or the court's 34-page order. It would be a material misrepresentation to present this as a case in which a regulator challenged an automated underwriting system, and nothing here does. The SEC filings are self-reported but are filed under securities liability, which is the reason they are used at all and the limit of what they establish.

• assumed

  NO PROTECTED-CLASS OR FAIR-LENDING EVIDENCE EXISTS IN THIS RECORD, AND NONE IS IMPORTED. Nothing anywhere in it alleges or measures disparate outcomes by race, sex, age, national origin or any other protected characteristic, and no fair-lending theory is pleaded. The pleading's population claim is economic: that the product is directed at consumers the operator considers financially vulnerable or financially coping. The figures that could be mistaken for disparities are shares of DETERMINATIONS rather than measurements on a subpopulation of served people, and every one of them is denied with the computation method itself in dispute. The parallel private class action raises a servicemember-specific statutory theory under the Military Lending Act, but it is a private action, is stayed pending a Ninth Circuit appeal, and is likewise unadjudicated. No node, edge, baseline, stressor or lever on this board rests on any of it.

• baseline

  THE SPEED ASYMMETRY IS THE GOVERNANCE FINDING, AND IT IS DRAWN RATHER THAN NARRATED. On one side: labelled repayment outcomes returning on an eleven-day cycle into a retraining loop, a publicly versioned engine, and an experiment apparatus the pleading alleges could measure an interface change's revenue effect within days. On the other: a civil investigative demand in January 2023, suit in November 2024, referral in December 2024, a dismissal ruling in September 2025, discovery through mid-2026 and a pretrial conference set for November 2026, with no merits ruling at any point. That is an order-of-magnitude difference in adaptation speed between the system and the check on it, and it is why the two retraining pathways are the widest on the board while both reviewer check arms sit at 1. Every figure on the fast side is the operator's own and unaudited; every date on the slow side is from the docket.

• baseline

  MOST OF THIS CASE IS NOT ABOUT THE AMOUNT, AND THE BOARD MUST NOT BE READ AS IF IT WERE. Three of the five counts are Restore Online Shoppers' Confidence Act counts about a monthly negative-option subscription charged to every consumer who linked a bank account, whether or not any advance was ever offered, and about the difficulty of stopping it; a fourth is about the tip mechanics. Only the first count concerns the advance representations. That is why the advertised-terms store carries so much of this diagram's traffic, why the support class is drawn with contact and without authority, and why the reconciliation pathway runs to the disclosure rather than to the engine. THE CONDUCT IS ALSO LARGELY HISTORIC: optional tips and express fees were removed for members onboarded from 4 December 2024 and eliminated in the February 2025 transition to a mandatory 5 per cent overdraft service fee with a $5 minimum, and the operator pleads mootness on that basis. Nothing here describes a tip slider, charitable-meal imagery or an express fee as a present feature. One detail is left unresolved rather than smoothed: the fiscal 2025 annual report gives a $3 monthly membership fee for new members from mid-2025 while the live site in August 2026 states an up-to-$5 monthly membership fee, one reporting period apart and one of them a ceiling rather than a rate, so no single current subscription price is asserted anywhere in this bundle.

• assumed

  WHERE THE RECORD IS SILENT, THE CONSERVATIVE VALUE, AND THE SILENCES ARE UNUSUALLY LARGE HERE. No source publishes an error rate, a defect rate, an override rate, a complaint-resolution rate, a reconciliation rate or a false-positive rate for any surface in this deployment — not the operator, not the government, not the court, and not any third party, because no third party has looked. The magnitudes on this board are therefore ordered by the structure of the record rather than by measurement, exactly as the PAN entry's own estimated-on-every-value discipline requires, and every one of PAN's parameters for this deployment carries `estimated: true`. Two consequences are stated plainly. First, the one quantity both parties argue about is a statistic ABOUT the engine's outputs rather than their accuracy, so this board's copy never asserts that the engine is accurate or inaccurate. Second, an absence of evidence is never scored as a deficit: the empty independent-evaluation channel is drawn as a pathway carrying nothing, which is what the public record supports, and not as a finding that the engine is unsound.

• assumed

  SERVED PEOPLE ARE NOT IN THE DYNAMICS. Members who link an account and receive an amount, members charged a monthly fee in a month in which the pleading alleges they were offered nothing, members shown a number and no reason for it, and members who tried to cancel are a boundary population recorded in the case file. No node, edge, baseline or lever here computes an approval, a fee, an overdraft, a household outcome or a harm for any person, and none is derived from this network. The pleaded consumer accounts — one consumer requiring 27 days and nine messages to support, which the court's order recites as 29 days — are carried in the case file as allegations and are never converted into a measurement on this diagram.

What this example does not show

• Everything the government says here is an ALLEGATION. The motion to dismiss was denied in full on 12 September 2025, which means the claims are plausible on the pleadings with the allegations taken as true. It does not mean anything was found, proved or established. The defendants deny paragraphs 34, 35 and 36 — the entire advance-distribution allegation — and deny liability generally.
• The advance-distribution figures are contested measurement twice over. The defendants contend the government's method of calculating the advances is wrong and its data incomplete, and the court expressly declined to resolve that dispute. The pleading also never dates the fourteen-month window the figures are drawn from. They appear here as the government's computation from the operator's own data, disputed by the operator, over an unspecified period.
• No regulator, court or auditor has examined the underwriting engine. Every statement about it — the named system, the training scale, the eleven-day refinement cadence, the version history, the claimed improvements in risk ranking and delinquency — comes from the operator's own SEC filings and press releases and is unaudited. This is not a case in which a regulator challenged an automated decision system.
• Most of this case is not about the amount. Three of the five counts concern a monthly negative-option subscription and the difficulty of cancelling it, and a fourth concerns the tip mechanics. Only one count concerns the advance representations.
• The conduct is largely historic. Optional tips and express fees were removed for members onboarded from 4 December 2024 and eliminated in the February 2025 transition to a mandatory 5 per cent overdraft service fee with a $5 minimum, and mootness is one of the operator's pleaded defences. This network describes no tip slider, no charitable-meal imagery and no express fee as present features. The current monthly subscription price is deliberately not asserted: the fiscal 2025 annual report and the August 2026 site state different things, one reporting period apart, and one of them is a ceiling rather than a rate.
• Nothing in this record alleges or measures disparate outcomes by race, sex, age, national origin or any other protected characteristic, and no fair-lending theory is pleaded anywhere. The pleading's population claim is economic. No fairness reading is available from this evidence and none is offered.
• No error rate, defect rate, override rate, complaint-resolution rate or false-positive rate is published for any surface in this deployment, by anyone. The magnitudes on this board are ordered by the structure of the record rather than by measurement, and the empty independent-evaluation pathway records what the public record contains, not a finding that the engine is unsound.
• Members are a boundary population. No approval, fee, overdraft, household outcome or harm for any person is computed here. The individual consumer accounts in the pleading — including the consumer alleged to have needed 27 days and nine messages to cancel, which the court's order recites as 29 days — are carried in the case file as allegations and are never turned into a measurement on this diagram.
• There is no settlement. Several low-quality sites assert a 2026 settlement, a claims portal or a payout date for this matter; none exists. The docket runs to 6 August 2026 with contested discovery and a final pretrial conference set for 9 November 2026, and the operator's Form 10-Q filed 5 August 2026 reports the matter unresolved with an aggregate legal-contingency accrual across its three pending consumer matters.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• In United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal.), the operative First Amended Complaint filed 30 December 2024 alleges that in the first fourteen months after Dave began advertising cash advances of 'up to $500', it offered a $500 advance to a new user about 0.002 per cent of the time — fewer than one determination in forty-five thousand; that only about 0.13 per cent of new users were offered even half of the advertised amount; that the most common offer, when an offer was made, was $25; that more than three-quarters of the time no advance was offered at all; and that on average more than 40 per cent of new users obtained no offer in a calendar month. Of new users who did receive offers, about 0.009 per cent of offers were for $500 and about 0.56 per cent were for at least $250; for existing users over the same window, on average more than a third were offered no advance in a calendar month and a $500 advance was offered less than 1 per cent of the time. THE DEFENDANTS DENY PARAGRAPHS 34, 35 AND 36 IN THEIR ENTIRETY and state in their dismissal brief that 'Dave can, does and did provide $500 advances'. The pleading never dates the fourteen-month window. Independently of the disputed tail, and undisputed because the operator publishes it, the advertised ceiling is $500 while Dave's own SEC-filed average advance was $170 in fiscal 2024 and $205 in fiscal 2025.

  empirical
  • Government First Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., filed 30 December 2024), ECF 44 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.44.0_1.pdf
  • Government Complaint for Permanent Injunction, Monetary Judgment, and Other Relief (unredacted, unsealed by court order), FTC v. Dave, Inc., No. 2:24-cv-09566 (C.D. Cal., filed 5 November 2024), ECF 24 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.24.0.pdf
  • Government Answer to Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 10 October 2025), ECF 80 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.80.0.pdf
  • Government Memorandum of Points and Authorities in Support of Defendants' Motion to Dismiss Plaintiff's Amended Complaint, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 28 February 2025), ECF 50-1 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.50.1.pdf
  • Government Dave Inc. (2026, March 2). Annual Report on Form 10-K for the fiscal year ended 31 December 2025, filed with the U.S. Securities and Exchange Commission https://www.sec.gov/Archives/edgar/data/1841408/000119312526085370/dave-20251231.htm

• The enforcement record says nothing whatever about the decision system. The words algorithm, artificial intelligence, machine learning, model, and underwriting appear nowhere in the original complaint, the operative amended complaint, the defendants' motion-to-dismiss memorandum, the government's opposition, the defendants' answer, or the court's 34-page order; the pleading describes only that Dave 'uses its access to consumers' bank accounts to analyze their finances and banking history' and 'uses this information to make decisions about how much (if any) to advance the consumer'. Every statement about the engine therefore comes from the operator's own investor-facing disclosures. Dave's Form 10-K for fiscal 2025 states that it uses 'our proprietary AI-powered underwriting system, CashAI' to 'analyze a Member's checking account transaction data to determine eligibility and set the bank's credit approval amount', in a 'fully automated process' that 'requires no credit check and does not rely on FICO or credit bureau data', drawing on 'hundreds of data points — including income patterns, spending behavior, and transaction history', and that the system has 'leveraged insights from over 180 million ExtraCash originations and billions of bank transactions'. The same filing discloses three further automated components in the same deployment: a real-time behavioural fraud-mitigation system with user-level controls, an income-and-expense prediction component feeding both underwriting decisions and the member-facing budgeting feature, and an automated support chatbot. In September 2025 the company announced CashAI v5.5, described as trained on more than 7 million recent originations that had reached full maturity, nearly doubling the prior feature set and optimized for the current fee structure, with claimed improvements in risk ranking, approval amounts, conversion, delinquency, and loss. ALL OF THAT IS THE OPERATOR'S OWN, UNAUDITED CLAIM. No regulator, court, or auditor has examined, described, or characterized the engine, and no model documentation, validation report, fairness assessment, or independent evaluation of it exists in the public record.

  empirical
  • Government Dave Inc. (2026, March 2). Annual Report on Form 10-K for the fiscal year ended 31 December 2025, filed with the U.S. Securities and Exchange Commission https://www.sec.gov/Archives/edgar/data/1841408/000119312526085370/dave-20251231.htm
  • Vendor Dave Inc. (2025, September 10). Dave Introduces CashAI v5.5 (press release) https://dave.com/press/news-releases/news-release-details/dave-introduces-cashai-v55
  • Government First Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., filed 30 December 2024), ECF 44 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.44.0_1.pdf
  • Government Memorandum of Points and Authorities in Support of Defendants' Motion to Dismiss Plaintiff's Amended Complaint, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 28 February 2025), ECF 50-1 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.50.1.pdf
  • Government Civil Minutes - General: (In Chambers) Order Denying Defendants' Motion to Dismiss, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 12 September 2025), ECF 75 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.75.0.pdf
  • Government Answer to Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 10 October 2025), ECF 80 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.80.0.pdf

• Three of the five counts are Restore Online Shoppers' Confidence Act counts and they concern the order in which things happen rather than the amount. The pleading alleges that a monthly membership fee was charged to every consumer who linked a bank account, whether or not any advance was ever offered, and that from at least August 2021 through November 2022 no in-app mechanism existed to stop that charge for consumers who also held a Dave bank account — which, from early 2022, Dave required of new consumers who wanted advances. It alleges at least nine separate in-app steps from the main screen to complete cancellation, diversion from cancellation for consumers who select the most prominent option, identity checks demanded to cancel including date of birth, sign-up phone number, mailing address, the last four digits of a Social Security number and details of the last two transactions on the external bank account, a July 2020 customer-service instruction that only consumers with no open advance and no pending advance payment were eligible to pause, and one consumer who required 27 days and nine messages to support and a threat to contact the Better Business Bureau (the court's order recites 29 days). A fourth count concerns the historic tip mechanic: a default charge of 15 per cent behind a large green 'Thank you!' button above imagery of a cartoon child and boxes reading '10 Healthy Meals', '15 Healthy Meals', '20 Healthy Meals', with the custom-tip alternative rendered white on white at about half the width and the child replaced by an empty plate at a zero tip; the pleading alleges Dave donated ten cents per percentage point of tip, usually $1.50 or less per advance, and kept the rest. Dave admits that tipping 'was formerly a revenue source', that members were 'presented with the option of providing an optional tip after the ExtraCash overdraft was sent', and that it 'donated a portion of each tip', and denies the remainder. The FTC's press release of 5 November 2024 states — citing Dave's own SEC filings and NOT the complaint — that Dave reported more than $149 million in revenue from these tips from 2022 through the first six months of 2024.

  empirical
  • Government First Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., filed 30 December 2024), ECF 44 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.44.0_1.pdf
  • Government Civil Minutes - General: (In Chambers) Order Denying Defendants' Motion to Dismiss, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 12 September 2025), ECF 75 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.75.0.pdf
  • Government Answer to Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 10 October 2025), ECF 80 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.80.0.pdf
  • Government Federal Trade Commission (2024, November 5). FTC Takes Action Against Online Cash Advance App Dave for Deceiving Consumers, Charging Undisclosed Fees; and (2024, December 30) FTC Refers Case Against Online Cash Advance Firm Dave Inc. to Department of Justice https://www.ftc.gov/news-events/news/press-releases/2024/11/ftc-takes-action-against-online-cash-advance-app-dave-deceiving-consumers-charging-undisclosed-fees

• The two loops in this deployment run at different orders of magnitude, and both speeds are documented. On the operator's side, Dave's Form 10-K for fiscal 2025 states that the approximately eleven-day average term of an advance 'creates rapid feedback loops, enabling iterative model refinement', and the company publicly versions the result. The pleading alleges a comparable apparatus on the interface: an experiment removing the 'Healthy Meals' content for some consumers, after which the percentage of new users charged a tip fell by about a third and overall tip revenue fell by almost a quarter, followed by an internal analysis recommending the content resume for all users; and a second experiment removing a three-box screen that likewise reduced both the number of consumers charged and the amounts. The pleading further alleges that dissatisfaction was measured with precision and answered without correction: an internal analysis of customer-service data naming 'Low advance amount', 'Low advance limits and approval', and 'Advance request denied' among the top drivers of contact; an internal survey naming 'Not enough money' a top source of dissatisfaction; thousands of monthly cancellation contacts of which 'most don't qualify for an advance or get a smaller than expected advance'; hundreds of monthly contacts on the topic 'What is the $1 charge?'; an internal analysis of Better Business Bureau complaints flagging 'inability to cancel easily within the app'; and an internal presentation stating that on the Express Fee screen 'what we promised is not what they see' and recommending Dave 'set expectations much earlier on the true cost of the money they are borrowing'. The pleading alleges each corresponding recommendation went unimplemented; the defendants refer the court to the documents in their entirety and deny mischaracterizations. On the other side, the consumer-protection loop ran on a civil investigative demand served in January 2023, suit in November 2024, referral in December 2024, a dismissal ruling in September 2025, contested discovery through mid-2026, and a final pretrial conference set for 9 November 2026, with no ruling on the merits at any point.

  empirical
  • Government First Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., filed 30 December 2024), ECF 44 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.44.0_1.pdf
  • Government Answer to Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 10 October 2025), ECF 80 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.80.0.pdf
  • Government Dave Inc. (2026, March 2). Annual Report on Form 10-K for the fiscal year ended 31 December 2025, filed with the U.S. Securities and Exchange Commission https://www.sec.gov/Archives/edgar/data/1841408/000119312526085370/dave-20251231.htm
  • Government Docket, Federal Trade Commission v. Dave, Inc., No. 2:24-cv-09566 (C.D. Cal.), entries 1-145 (5 November 2024 to 6 August 2026), CourtListener / RECAP Archive https://www.courtlistener.com/docket/69345740/federal-trade-commission-v-dave-inc/
  • Government Federal Trade Commission (2024, November 5). FTC Takes Action Against Online Cash Advance App Dave for Deceiving Consumers, Charging Undisclosed Fees; and (2024, December 30) FTC Refers Case Against Online Cash Advance Firm Dave Inc. to Department of Justice https://www.ftc.gov/news-events/news/press-releases/2024/11/ftc-takes-action-against-online-cash-advance-app-dave-deceiving-consumers-charging-undisclosed-fees

• The disclosure moved while the litigation ran, and the before-and-after is directly observable without discovery. The amended complaint at paragraph 22 quotes Dave's website shortly after the November 2024 filing as carrying 'Get up to $500 in 5 minutes or less' with a fine-print footnote stating only that 'the average advance is $170' and that 'enrollment and initial qualification [are] typically completed in 5 minutes', and alleges that even that footnote failed to disclose that many consumers who give Dave bank-account access will be offered no advance at all. The same site as displayed on 28 August 2026 still leads with 'Up to $500 in 5 min or less', and its footnote now reads: 'ExtraCash amounts range from $25-$500, typically authorized within 5 minutes, with an overdraft fee equal to the greater of $5 or 5%. Multiple overdrafts may be required. Not all members qualify for ExtraCash and few qualify for $500.' The two concessions now present — that not all members qualify at all, and that few qualify for $500 — are precisely the two omissions the amended complaint pleads. The price surfaces moved too, while liability was denied: members onboarded from 4 December 2024 were placed on a structure without optional tips or express fees, and in February 2025 Dave completed a transition to a mandatory 5 per cent overdraft service fee with a $5 minimum, with tip revenue falling 89 per cent from $67.6 million in fiscal 2024 to $7.5 million in fiscal 2025 and subscription revenue rising 51 per cent to $37.2 million. ONE DETAIL IS DELIBERATELY LEFT UNRESOLVED: the fiscal 2025 annual report reports a $3 monthly membership fee for new members from mid-2025 while the live site in August 2026 states an 'Up to $5 monthly membership fee', one reporting period apart and one of them a ceiling rather than a rate, so no single current subscription price is asserted.

  empirical
  • Government First Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc. and Jason Wilk, No. 2:24-cv-09566-MRA-AGR (C.D. Cal., filed 30 December 2024), ECF 44 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.44.0_1.pdf
  • Vendor Dave Inc. (2026). Consumer marketing site, homepage and ExtraCash disclosures as displayed 28 August 2026 https://dave.com/
  • Government Dave Inc. (2026, March 2). Annual Report on Form 10-K for the fiscal year ended 31 December 2025, filed with the U.S. Securities and Exchange Commission https://www.sec.gov/Archives/edgar/data/1841408/000119312526085370/dave-20251231.htm
  • Trade press Banking Dive (2025). Fintech Dave alters fee structure, blasts DOJ over amended lawsuit https://www.bankingdive.com/news/fintech-dave-alters-fee-structure-blasts-doj-over-amended-lawsuit-jason-wilk/736830/

• On 12 September 2025 Judge Monica Ramirez Almadani denied the motion to dismiss in full in a 34-page order, holding that 'numerous courts have found that "up to" representations can materially mislead reasonable consumers where the defendant does not or cannot provide the good or service as represented, especially when the representation references a particular, quantified amount', that 'the government has plausibly alleged that it was exceedingly rare for Dave to offer the maximum amount of the cash advance advertised or even amounts approaching the maximum', and that a fine-print 'Terms apply' disclaimer in two banner advertisements did not cure the net impression because 'a disclaimer does not automatically exonerate deceptive activities'. THAT IS A PLAUSIBILITY RULING ON THE PLEADINGS AND ESTABLISHES NOTHING FACTUAL, and the order's footnote 2 records the dispute it did not resolve: 'Defendants contend that the government's method of calculating cash advances is wrong and that the data it used is [in]complete. ... Such a factual dispute cannot be resolved on a Rule 12(b)(6) motion to dismiss.' Dave and Wilk answered on 10 October 2025 with a general denial and seven affirmative defences: lack of fair notice of the government's interpretation of ROSCA; that ROSCA is unconstitutionally vague as applied; standing and mootness; good faith, resting in part on the assertion that 'the Consumer Financial Protection Bureau opened and closed an investigation — and declined to recommend an enforcement action against Dave' (an assertion by the defendant, with no agency document confirming it located); offsets; the statute of limitations; and that the penalties sought are unconstitutionally excessive. Publicly the company called the amended complaint 'a continued example of government overreach' resting on 'numerous allegations that are based on various inaccuracies', said it believes it has 'always acted within the law', and pledged to 'vigorously defend itself'. A Civil Trial Order of 14 November 2025 set a final pretrial conference for 9 November 2026; a stipulated protective order was entered in January 2026; contested discovery ran from March through June 2026; the last docket activity as of 6 August 2026 is counsel withdrawals. THERE IS NO SETTLEMENT, NO CONSENT ORDER AND NO ADJUDICATION ON THE MERITS, and the operator's Form 10-Q filed 5 August 2026 states it is 'unable to reasonably predict the possible outcome' and records a $9.7 million aggregate legal-contingency accrual across its three pending consumer matters. The 2025 change of FTC leadership did not thin the case: both authorizing Commission votes were 4-1 with Commissioner Melissa Holyoak voting no, the United States became the real party in interest in December 2024, and the Department of Justice litigated the matter through 2025 and 2026. Two private actions run alongside and are likewise unadjudicated: a Military Lending Act and Truth in Lending Act putative class action naming Dave and Evolve Bank & Trust, in which dismissal and arbitration were both denied on 12 December 2025 and which is on appeal to the Ninth Circuit with district proceedings stayed, and a suit by the Mayor and City Council of Baltimore under a municipal consumer-protection ordinance.

  empirical
  • Government Civil Minutes - General: (In Chambers) Order Denying Defendants' Motion to Dismiss, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 12 September 2025), ECF 75 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.75.0.pdf
  • Government Answer to Amended Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief, United States v. Dave, Inc., No. 2:24-cv-09566-MRA-AGR (C.D. Cal., 10 October 2025), ECF 80 https://storage.courtlistener.com/recap/gov.uscourts.cacd.947304/gov.uscourts.cacd.947304.80.0.pdf
  • Government Docket, Federal Trade Commission v. Dave, Inc., No. 2:24-cv-09566 (C.D. Cal.), entries 1-145 (5 November 2024 to 6 August 2026), CourtListener / RECAP Archive https://www.courtlistener.com/docket/69345740/federal-trade-commission-v-dave-inc/
  • Government Dave Inc. (2026, August 5). Quarterly Report on Form 10-Q for the quarterly period ended 30 June 2026, filed with the U.S. Securities and Exchange Commission https://www.sec.gov/Archives/edgar/data/1841408/000119312526335154/dave-20260630.htm
  • Government Federal Trade Commission (2024, November 5). FTC Takes Action Against Online Cash Advance App Dave for Deceiving Consumers, Charging Undisclosed Fees; and (2024, December 30) FTC Refers Case Against Online Cash Advance Firm Dave Inc. to Department of Justice https://www.ftc.gov/news-events/news/press-releases/2024/11/ftc-takes-action-against-online-cash-advance-app-dave-deceiving-consumers-charging-undisclosed-fees
  • Trade press Banking Dive (2025). Fintech Dave alters fee structure, blasts DOJ over amended lawsuit https://www.bankingdive.com/news/fintech-dave-alters-fee-structure-blasts-doj-over-amended-lawsuit-jason-wilk/736830/

Where this connects

Institutional pressures in this domain

• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Lending & credit collections AI domain page.

Levers available here and the patterns behind them

• Upgrade model — Improve the model
• Check with a second model — Cross-model verification
• Check copied records — Reconcile copied records
• Understand the system — Understand the system
• Gate record entries — Human-in-the-loop write gating
• Verify output — Put a verifier on the agent
• Assign a challenger — Structured dissent
• Keep prompts neutral — Framing and mirroring reduction
• Review on schedule — Oversight cadence & retrospectives
• Store less data — Data minimization
• Pause AI on alarms — Deployment circuit-breaker

Documented case histories

• Dave ExtraCash: an advertised ceiling, an automated amount, and a case that never asks how the amount is set
• Automated underwriting with its fair-lending testing on the record
• Cleared on the numbers but faulted on the explanation
• The governance an enforcement action had to write
• M-Shwari & Kenya's Digital Credit Market
• Citi Retail Services Judgmental Review & the Armenian surname screen
• Santander Consumer USA subprime vehicle loan scoring
• Credit Acceptance Corporation's net-collections score
• Wells Fargo refinance underwriting & the bridge nobody could build
• Navy Federal mortgage underwriting & three readings of one gap
• Enova International servicing defects & the debits nobody authorised
• Equifax Online Model Server coding error (2022)
• TransUnion's OFAC Name Screen & the people who could not sue
• Hello Digit's automated-savings algorithm
• Oportun's legal-collections filing pipeline