Inside the Upstart Lawsuit: What the Complaint Actually Says About Model 22

On April 7, 2026, plaintiff Anthony Dunn filed a 25-page class action complaint against Upstart Holdings, Inc. and four of its senior executives in the US District Court for the Northern District of California. The lawsuit — Dunn v. Upstart Holdings, Inc., Case No. 3:26-cv-02974-JD — has been covered in the trade press, but the full complaint tells a more specific story than the headlines suggest.

It is worth reading carefully, because the legal theory it constructs has implications well beyond Upstart.

The timeline the complaint establishes

The class period runs from May 14, 2025 to November 4, 2025. During that window, investors allege that Upstart executives made materially false and misleading statements about the accuracy and performance of Model 22, the company’s latest AI underwriting model launched in early May 2025.

The sequence of events the complaint documents is precise:

February 2025: Upstart issues full-year 2025 revenue guidance of approximately $1 billion, including roughly $920 million in fee revenue.

May 14, 2025: Upstart holds its inaugural “AI Day” investor event. CTO Paul Gu presents the technical architecture of Model 22, touting its use of neural networks at every level of the model architecture — a claimed 17 percentage point improvement in separation accuracy over the benchmark credit model to 171.2%. CEO Dave Girouard states that four goals for 2025 include “10X Upstart’s advantage in AI.” The company slightly raises revenue guidance to approximately $1.01 billion.

August 5, 2025: Upstart reports strong Q2 results and significantly raises full-year guidance to approximately $1.055 billion, with fee revenue guidance increased by $70 million to $990 million. Girouard attributes the performance “first and foremost” to Model 22. CFO Sanjay Datta states that transactional revenue “more than doubled year-on-year, largely reflecting the influence of the aforementioned Model 22.”

November 4, 2025: Upstart reports Q3 revenue of $277 million, missing guidance by approximately $3 million. Q4 guidance comes in at $288 million, significantly below consensus of $303.7 million. Full-year guidance is cut to $1.035 billion from $1.055 billion. Fee revenue guidance falls to $946 million from $990 million — a $44 million downward revision.

The stock falls 9.71% the following day, closing at $41.75 per share.

What executives said on the November earnings call

This is where the complaint gets most specific — and most damaging. The plaintiffs quote directly from the Q3 2025 earnings call transcripts, and the language executives used is notable.

Girouard, in prepared remarks: “Our risk models responded to macroeconomic signals they observed by moderately reducing approvals and increasing interest rates. This drove a reduction in our conversion rate from 23.9% in Q2 to 20.6% in Q3.”

When an analyst asked how to reconcile strong application demand with the revenue miss, Girouard attributed it to Model 22’s tendency to respond too quickly to macro signals, describing the model as having potentially been “overreacting” — while also saying “having a model that overreacts is better than having ones that underreact.”

Gu, in response to analyst questions, acknowledged that Upstart had “knowingly” calibrated their AI model to be “more conservative on the credit side in earlier parts of the quarter.” He also acknowledged the model was “a little overly responsive to the latest changes” and that there was “natural statistical sampling error” in the model’s calibration — describing work done to reduce that measurement error by approximately half.

Datta acknowledged that “the model impact in Q3, even though it appears to be abating, will impact Q4 as well.”

The legal theory — and why it matters beyond the stock drop

The complaint asserts two counts: violations of Section 10(b) of the Securities Exchange Act and Rule 10b-5 (the core securities fraud provision), and violations of Section 20(a) (control person liability against the individual defendants).

The core allegation is that executives knew — or recklessly disregarded — that Model 22 was prone to overreacting to macroeconomic signals, that this was already negatively impacting revenue during the class period, and that they failed to disclose this while raising guidance and touting the model’s accuracy. The complaint also cites SEC Regulation S-K Item 303, which requires companies to disclose known trends or uncertainties likely to have a material unfavorable impact on revenues. The investors argue that Model 22’s overresponsiveness was exactly such a known trend — one that was being concealed rather than disclosed.

The scienter allegation — the intent element required for securities fraud — rests heavily on insider stock sales. During the class period, Girouard sold 208,335 shares for proceeds of over $13.5 million. Datta sold 26,985 shares for over $1.4 million. Gu sold 5,000 shares for over $344,000. The complaint argues that executives who sold millions in stock while concealing known model deficiencies had both motive and opportunity to commit fraud.

What this means for AI-native lenders and their boards

The Upstart complaint is the first securities fraud lawsuit in the US consumer lending space to center specifically on an AI model’s behavior as the basis for securities misrepresentation claims. That makes it a template — and a warning — for any lender whose public narrative is built substantially around AI underwriting capabilities.

The disclosure obligation question is now live. When an AI model is exhibiting known behavioral characteristics — overresponsiveness to macro signals, measurement error, systematic bias in any direction — at what point does that create a disclosure obligation under Item 303? The Upstart complaint argues that the answer is: when it is already materially impacting revenue, and executives know it. This is a question every AI-native lender’s legal and compliance team should now be answering explicitly.

Guidance raised on the back of AI model performance is high-risk territory. Upstart raised full-year revenue guidance by $70 million in fee revenue specifically citing Model 22’s performance improvements. When that model subsequently underperformed, the delta between raised guidance and actual results became the evidentiary core of the lawsuit. Lenders whose forward guidance is linked to AI model performance assumptions need to be precise about what those assumptions are — and what happens if the model behaves differently than expected.

“Knowingly” is a dangerous word on an earnings call. Gu’s acknowledgment that Upstart had “knowingly” chosen to make the model more conservative appears multiple times in the complaint. In securities law, scienter — the intent to deceive — is one of the hardest elements to prove. The word “knowingly” coming from the CTO’s mouth during a public earnings call, in the context of a prior undisclosed model change, is exactly the kind of evidence plaintiffs’ lawyers spend months trying to find in discovery. It was handed to them in a transcript.

The leadership transition adds a complicating layer. Paul Gu — named as a defendant in his capacity as CTO — is scheduled to become Upstart’s CEO on May 1, 2026. He is inheriting a securities fraud lawsuit centered substantially on statements he made and model decisions he was responsible for. How that shapes Upstart’s litigation posture, settlement calculus, and investor communications going forward will be worth watching.

The case is at the complaint stage. Upstart has not yet filed a response. Class certification has not been granted. But the facts in the complaint — drawn directly from Upstart’s own earnings call transcripts, SEC filings, and press releases — are not in dispute. What is in dispute is whether they add up to securities fraud. That question will take years to resolve. The implications for how AI-driven lenders communicate model risk are already here.